[{"data":1,"prerenderedAt":10},["ShallowReactive",2],{"article-anthropic-2t-ipo-safety-trust-vs-public-shareholders":3},{"slug":4,"title":5,"summary":6,"date":7,"published":8,"content":9},"anthropic-2t-ipo-safety-trust-vs-public-shareholders","Anthropic's $2T IPO will test whether a self-appointed safety trust can survive public shareholders","Anthropic's planned IPO could value the company at $2 trillion and will expose its Long-Term Benefit Trust, a three-person body with no equity but majority board-appointment power, to public-market scrutiny for the first time. The trust has never drawn a red line or forced a profit-vs-safety trade-off. Its only constraint is an 85% shareholder kill switch that may dilute once retail investors enter the cap table. Harvard Law calls it a built-in conflict hard-wired into the corporate DNA. The Ben & Jerry's precedent says the conflict resolves against the guardians. The question is whether a safety trust designed for a private lab can survive a public one.","2026-09-09",true,"\u003Ch1>Anthropic's $2T IPO will test whether a self-appointed safety trust can survive public shareholders\u003C/h1>\n\u003Cp>In June 2026, we wrote that AI agents are hitting a permissions wall before they hit a model wall: the defining challenge is not whether the model is smart enough to be useful, but whether the governance infrastructure around it can contain what it does when it is dangerous. Three months later, Anthropic is about to test that thesis at the corporate level. The company that built Claude is preparing an IPO that could value it at $2 trillion, and the governance structure it designed to keep its AI safe, the Long-Term Benefit Trust (LTBT), is about to face the one test it has never taken: a conflict it cannot resolve by advising.\u003C/p>\n\u003Ch2>The structure: three people, no equity, majority board control\u003C/h2>\n\u003Cp>The LTBT is the controlling governance mechanism of Anthropic's public benefit corporation. It holds no equity in the company. It cannot sell shares, collect dividends, or profit from Anthropic's commercial success. What it can do is appoint and dismiss the majority of Anthropic's board of directors. It has selected four of the company's seven current directors, including Netflix co-founder Reed Hastings and Vas Narasimhan, the chief executive of Novartis.\u003C/p>\n\u003Cp>The trust currently has three members out of a potential maximum of five. It is chaired by Neil Buddy Shah, chief executive of the Clinton Health Access Initiative. The other two members are Ben Bernanke, the former Federal Reserve chair who joined on July 9, 2026, and Richard Fontaine, CEO of the Center for a New American Security. A fourth member, Mariano-Florentino Cuéllar, a former California Supreme Court justice, left the trust on August 4, 2026, not to return to academia but to become Anthropic's first chief global affairs officer. The move from overseer to executive is a detail critics of the structure will not let pass quietly.\u003C/p>\n\u003Cp>The trustees receive advance notice of major company actions, including the launch of new AI models. They meet weekly among themselves and as frequently as every other week with Anthropic's leadership. They attend regular board meetings and hold discussions with the founders on significant issues, including the Mythos cybersecurity model, where trustees encouraged a limited rollout through the Glasswing Project, and the company's dispute with the US government over automated weapons.\u003C/p>\n\u003Cp>That is significant access. It is also, so far, purely advisory.\u003C/p>\n\u003Ch2>The untested guardrail: advisory-only, zero red lines\u003C/h2>\n\u003Cp>Despite that access and authority, the trust has operated largely in an advisory capacity. It has not attempted to draw red lines or force a significant trade-off between profit and purpose. No one close to the trust's workings has described a moment where it told Anthropic's leadership to sacrifice revenue for safety, and was obeyed because it had the power to remove the board.\u003C/p>\n\u003Cp>This is the central problem, and Harvard Law professor Jesse Fried named it directly. Anthropic "raises funds from profit-seeking investors, then lets self-appointed individuals decide how much profit to sacrifice for the firm's mission," Fried wrote in a July 2026 paper. "A deep and potentially unmanageable tension is thus hard-wired into the firms' corporate DNA."\u003C/p>\n\u003Cp>Elizabeth Pollman, a corporate governance expert at the University of Pennsylvania, put the structural challenge more cautiously: "It's nearly impossible to perfectly contract for all possible circumstances that could arise when managing competing interests within a firm." She pointed to the intensity of AI competition at the geopolitical and corporate levels as factors that make the balancing act harder, not easier. "Will this governance structure work in the way intended, serving dual or more interests over time? That's the real challenge."\u003C/p>\n\u003Cp>The honest read is that the LTBT is a governance experiment that has been running in a lab with no stress applied. The IPO is the stress test.\u003C/p>\n\u003Ch2>The kill switch: 85%, and the math that makes it fragile\u003C/h2>\n\u003Cp>Anthropic's trust is considered less risky than OpenAI's governance structure because it has a built-in kill switch: the trustees can be dismissed with the support of 85% of shareholders' voting power. Fried considers this constraint the key feature that makes the arrangement workable. "It puts a constraint on the guardians, because they don't want to be thrown out," he told Fortune. "If investors really don't like what the guardians are planning to do, I expect the guardians will back off."\u003C/p>\n\u003Cp>That logic held in private markets. The investors who backed Anthropic across nine primary rounds (Amazon at roughly 21%, Alphabet at roughly 15%, and the rest a roster of aligned venture capitalists) did so with full knowledge of the governance structure. Several specifically cited Anthropic's safety emphasis as part of their investment thesis. These are investors who bought into the mission before they bought into the multiple.\u003C/p>\n\u003Cp>But the 85% threshold is a supermajority that could change when the company goes public, according to a person with knowledge of the structure. Early investors bought in knowing the governance bargain. Retail investors will not. A supermajority that was achievable among a concentrated group of aligned VCs becomes a referendum among diffuse shareholders who bought the stock because it was the largest tech IPO in history, not because they read the trust charter.\u003C/p>\n\u003Cp>The mechanical question is whether the kill switch constrains the trustees or constrains the shareholders. If 85% of a public float can be assembled to fire the trust the first time it tries to slow a product launch, the trust is not a guardrail; it is a suggestion. And if the threshold is too high to reach in practice, the trust is unaccountable to the very public investors funding the company. Either way, the kill switch is a feature that works best in the market structure that created it, not the one it is entering.\u003C/p>\n\u003Ch2>The Ben & Jerry's precedent: when guardians backfire\u003C/h2>\n\u003Cp>Fried and Reiter's Harvard Law paper does not stop at naming the conflict. It identifies a pattern they call "Ben & Jerry's risk": the danger that mission guardians will not only harm investors but achieve the exact opposite of what they set out to do.\u003C/p>\n\u003Cp>When Unilever acquired Ben & Jerry's in 2000, it installed self-perpetuating independent directors who could override the parent company to protect the brand's social mission. For two decades, tensions stayed behind closed doors. Then in 2021, the independent board refused to renew the Israeli licensee's license over Unilever's objections. Counterboycotts, state divestments, shareholder lawsuits, and the resignation of Unilever's CEO followed. Unilever's market cap dropped $20 to $26 billion in the months after. Seven states divested pension fund holdings totaling nearly $1 billion. Unilever then overrode the directors, gave the licensee perpetual rights (exactly what the guardians tried to prevent), and spun off its ice cream businesses altogether, ensuring the guardians could never impose costs on the parent again.\u003C/p>\n\u003Cp>Fried's assessment is blunt: "Basically these people could do whatever they wanted, no matter how much damage it would inflict on Unilever, and they couldn't easily be removed." He called it "an ill-considered arrangement" that he assumed no one would ever replicate. Enter Anthropic.\u003C/p>\n\u003Cp>The parallel is not exact. Ben & Jerry's guardians were protecting a social mission in a consumer goods company; Anthropic's trustees are overseeing AI safety in a company whose product can rewrite infrastructure. The stakes are higher. But the structural failure mode is the same: guardians with no economic interest, accountable to a charter rather than to shareholders, making decisions that impose costs on the people who funded the enterprise. When those costs get large enough, the enterprise removes the guardians. The mission the guardians were protecting does not survive the removal.\u003C/p>\n\u003Cp>OpenAI's November 2023 board firing is the AI-specific version of the same pattern. The board tried to fire Sam Altman over safety concerns, lost the confidence of investors and employees, and was replaced. Altman returned, the safety-focused directors were pushed out, and prominent safety researchers departed to start competing ventures. Fried's read: "They not only put investors at risk but achieved the exact opposite of their mission, as they saw it. They thought that Sam Altman could not be trusted to lead a safe OpenAI. He's still there. All the board members who cared about safety are gone."\u003C/p>\n\u003Ch2>The GAAP ambition: a blueprint or a liability shield?\u003C/h2>\n\u003Cp>Anthropic's ambition for the LTBT extends beyond its own corporate structure. Multiple people close to the company's plans say Anthropic is attempting to establish a template for future AI governance through the trust: an industry blueprint similar to GAAP accounting principles, which initially emerged as voluntary standards proposed by private companies. Bernanke's appointment in July is seen as a signal of this institutional ambition.\u003C/p>\n\u003Cp>This is the tell. The GAAP comparison is aspirational, but GAAP works because it was built by an independent standards body (the Financial Accounting Standards Board), not by the company whose financials are being measured. A governance blueprint designed by the lab whose AI it is supposed to oversee is not an independent standard; it is a corporate governance structure that the lab chose for itself and is now asking the industry to adopt.\u003C/p>\n\u003Cp>This connects to a broader pattern this blog has tracked: AI governance frameworks written with the labs' input tend to serve the labs' interests. The White House frontier-licensing process, where OpenAI's models are approved customer-by-customer with no published rubric, is a regulatory framework shaped by the labs it regulates. The LTBT is a corporate governance framework shaped by the lab it governs. In both cases, the entity being regulated designed the regulator. The difference is that the White House process is a political arrangement with no legal teeth; the LTBT is a corporate structure with real board-removal power, which makes it more consequential and more dangerous if it fails.\u003C/p>\n\u003Cp>The right question is not whether the LTBT is a thoughtful structure. It is. The right question is whether a safety structure designed by the lab, staffed by the lab's appointees, and overrideable by the lab's investors can ever be more than a liability shield dressed as governance. The Ben & Jerry's precedent says no. The OpenAI precedent says no. The LTBT has not yet been tested, and an IPO that values the company at $2 trillion is the test.\u003C/p>\n\u003Ch2>What investors, enterprises, and policymakers should watch\u003C/h2>\n\u003Cul>\n\u003Cli>\n\u003Cp>\u003Cstrong>Read the S-1 risk factor on the trust's powers and any sunset provisions.\u003C/strong> The prospectus went public in late September, and the document that matters is not the valuation range: it is the risk factor describing the LTBT's authority, any sunset or expiration language, and whether the 85% kill switch survives the IPO unchanged. That language determines whether large index funds and pension mandates can hold the stock at weight. If the powers never expire, some index providers will restrict the stock from major benchmarks, constraining the float.\u003C/p>\n\u003C/li>\n\u003Cli>\n\u003Cp>\u003Cstrong>Watch for the first profit-vs-safety trade-off.\u003C/strong> The trust has never drawn a red line. The first time it does (slowing a model launch, blocking a deployment, overriding a commercial decision) will be the moment the structure is actually tested. If the trustees back down, the trust is advisory. If they hold and the shareholders fire them, the trust is the Ben & Jerry's board. Either outcome validates the critics.\u003C/p>\n\u003C/li>\n\u003Cli>\n\u003Cp>\u003Cstrong>Demand the trust's decision logs as a condition of enterprise contracts.\u003C/strong> Anthropic's enterprise customers buy Claude partly because of the company's safety posture. The trust is what makes that posture credible. Enterprises should require, as a contractual condition, access to the trust's meeting minutes and decision records, not Anthropic's summaries but the primary documents. A trust that operates behind closed doors is a marketing claim, not a governance mechanism.\u003C/p>\n\u003C/li>\n\u003Cli>\n\u003Cp>\u003Cstrong>Treat the GAAP ambition with skepticism.\u003C/strong> Anthropic wants the LTBT to become an industry blueprint. Enterprises and regulators should evaluate whether a governance standard designed by the lab it oversees meets the independence bar of standards they already rely on. GAAP was not designed by the companies whose financials it measures. An AI governance blueprint designed by an AI lab is not the same thing.\u003C/p>\n\u003C/li>\n\u003C/ul>\n\u003Ch2>The bet\u003C/h2>\n\u003Cp>Anthropic's private investors assumed capitalism would win in the end. As one venture capitalist who backed the company put it: "Whatever you say, if you need a lot of money for compute and to compete for the best model, investors assume there will be a business." The LTBT was designed for a world where that assumption could coexist with safety oversight, where the trustees could advise, the company could grow, and the tension between profit and purpose would never reach a breaking point.\u003C/p>\n\u003Cp>The IPO exposes that assumption. A $2 trillion valuation requires Anthropic to deliver roughly $190 to $200 billion in 2028 revenue, roughly tripling from the July 2026 run rate of $65 billion. That growth path requires aggressive product launches, enterprise expansion, and the kind of competitive deployment velocity that creates exactly the profit-vs-safety trade-offs the trust was built to arbitrate. The trust has never had to arbitrate one. The IPO is the moment it starts having to.\u003C/p>\n\u003Cp>The structural bet is this: the LTBT can survive public shareholders only if it never has to use its power. The moment it does, the first time it tells Anthropic to slow a deployment and the market disagrees, the 85% kill switch becomes the mechanism by which the safety mission is removed. A trust that cannot survive its own first test is not a guardrail. It is a marketing asset that the IPO prospectus will describe as governance and the market will price as risk.\u003C/p>\n\u003Cp>Anthropic is the most safety-credible lab in the market. Its governance structure is more thoughtful than OpenAI's. Its trustees are serious people. None of that changes the structural fact: a safety trust designed for a private company is being tested by the market that private companies go public to access, and the market's tolerance for safety-induced friction is lower than any private investor's has been. The IPO is not just a valuation event. It is the first real stress test of whether mission-driven AI governance can survive the market it was designed to operate above.\u003C/p>\n\u003Chr>\n\u003Cp>\u003Cem>Sources: \u003Ca href=\"https://arstechnica.com/ai/2026/09/anthropics-2-trillion-ipo-puts-powerful-external-trustees-in-spotlight/\">Ars Technica: "Anthropic's $2 trillion IPO puts powerful external trustees in spotlight"\u003C/a> (Financial Times syndication, September 2026); \u003Ca href=\"https://fortune.com/2026/06/01/openais-guardian-ben-jerrys-ice-cream-anthropic/\">Fortune: "Harvard Law: Anthropic is about to sell a safety mission Wall Street can veto"\u003C/a> (Catherina Gioino, June 2026); \u003Ca href=\"https://valueaddvc.com/pulse/anthropic-ipo-long-term-benefit-trust-trustees-2026\">Value Add Pulse: "Three Outsiders Will Control Anthropic's Board After IPO"\u003C/a> (September 2026); \u003Ca href=\"https://graniteshares.com/research/anthropic-ipo-2026-explained-from-965-billion-to-a-possible-2-trillion-listing/\">GraniteShares: "Anthropic IPO 2026 Explained"\u003C/a> (August 2026); \u003Ca href=\"https://www.reuters.com/business/finance/anthropic-warns-ai-may-pose-existential-risks-humanity-ipo-filing-2026-09-29/\">Reuters: "Anthropic warns AI may pose 'existential risks to humanity' in IPO filing"\u003C/a> (September 28, 2026); \u003Ca href=\"https://corpgov.law.harvard.edu/2026/05/11/ai-corporate-governance-and-ben-jerrys-risk/\">Harvard Law: Jesse Fried & Idan Reiter, "AI Corporate Governance and Ben & Jerry's Risk"\u003C/a> (May 2026). Prior analysis: \u003Ca href=\"https://ivmanto.com/blog/ai-agents-permissions-wall\">AI agents and the permissions wall\u003C/a> (June 2026).\u003C/em>\u003C/p>\n",1791046633229]