Corporate Governance

A board that used its authority lasted five days

In November 2023 OpenAI's nonprofit board removed its chief executive, and was itself removed within five days. The restructuring completed in October 2025 gives a safety committee more formal power than that board ever held.

P J Laszkowicz

On 28 October 2025, OpenAI completed the most significant corporate restructuring in the history of artificial intelligence. The company that began as a nonprofit research laboratory dedicated to ensuring artificial general intelligence "benefits all of humanity" became a for-profit public benefit corporation valued at approximately $500 billion. The nonprofit that founded it, renamed the OpenAI Foundation, holds 26 per cent of the new entity, worth roughly $130 billion, which makes it one of the best-resourced philanthropic organisations in the world.

The restructuring took years. It survived a boardroom coup, weathered scrutiny from two state attorneys general, and prompted one of the most consequential debates in technology governance: what happens when the institution built to steward the most powerful technology in human history decides it must become a conventional corporation in order to do so?

The question here is not whether OpenAI made the right decision. It is what the decision reveals about the structural tension between mission-driven governance and the capital demands of frontier AI, and about whether any corporate form, however carefully designed, can hold that tension when billions of dollars and geopolitical ambition are at stake.

From research lab to $500 billion company

The path from nonprofit laboratory to half-trillion-dollar corporation took exactly ten years. Each structural change answered the same problem: the mission required resources the current structure could not attract.

OpenAI was founded in December 2015 as a nonprofit research organisation.1 Its co-founders included Sam Altman, then president of the startup accelerator Y Combinator, and Elon Musk, alongside Ilya Sutskever, Greg Brockman, and several other researchers and entrepreneurs.2 Donors including Altman, Musk, Reid Hoffman, Peter Thiel, and Amazon Web Services pledged $1 billion.3 The mission was to develop artificial general intelligence safely and ensure its benefits were widely shared. The nonprofit structure was the point: no investor returns, no shareholder pressure, no duty owed to anyone but the mission.

The first structural crisis came within four years. By 2018 it was clear that training competitive models required computing resources a nonprofit could not sustain. In March 2019 OpenAI created a capped-profit subsidiary, OpenAI LP, controlled by the nonprofit parent.4 Investors could earn returns capped at one hundred times their investment, a figure that seemed generous until the company's valuation made the cap irrelevant.5 The structure preserved the nonprofit board's ultimate authority. OpenAI LP's governing documents stated that if the board concluded the organisation's activities were inconsistent with its mission, it could redirect the company, up to and including dissolving it.

Microsoft entered in July 2019 with $1 billion, becoming OpenAI's exclusive cloud partner through Azure.6 In January 2023 it committed a further $10 billion in what it called the third phase of the partnership.7 By then Microsoft had invested approximately $13.8 billion in total.8 The arrangement was unusual: exclusive distribution rights for the programming interface and cloud obligations, but no seat on the nonprofit board and no formal governance role.

Then, on 17 November 2023, the structure was tested and it broke.

The five days

OpenAI's board, comprising chief scientist Ilya Sutskever, Quora chief executive Adam D'Angelo, entrepreneur Tasha McCauley, and Georgetown researcher Helen Toner, removed Sam Altman as chief executive.9 Its terse statement said only that Altman "was not consistently candid in his communications with the board".10

Five days of corporate chaos followed, exposing every tension the nonprofit structure was meant to manage. Microsoft, which had invested billions and held no board representation, was reportedly told minutes before the announcement.11 More than 700 of OpenAI's 770 employees signed a letter threatening to resign and join Microsoft unless Altman was reinstated.12 Sutskever reversed himself publicly on 20 November, saying he deeply regretted his part in the firing.13

On 21 November OpenAI announced Altman's return. The price was a reconstructed board: Bret Taylor, former co-chief executive of Salesforce, as chair; the former Treasury Secretary Larry Summers; and D'Angelo as the only holdover.14 McCauley, Toner, and Sutskever were gone. The board that had exercised its governance authority, the mechanism the nonprofit structure existed to enable, was itself removed within days.

Sutskever, the architect of the firing and one of the company's most important researchers, departed quietly in May 2024.15

The nonprofit board held the formal authority to remove the chief executive. It used that authority. What followed was not a governance correction but a governance collapse. The board that put safety concerns above commercial momentum was replaced by one that would not repeat the experiment.

The restructuring

Two years of negotiation followed over what OpenAI should become. In late 2024 the company proposed converting to a standard for-profit corporation and removing the nonprofit's controlling role entirely.16 The backlash was immediate. Musk sued. The attorneys general of California and Delaware, the states where OpenAI is headquartered and incorporated, signalled they would scrutinise any conversion stripping the nonprofit of its assets and mission.17 A coalition called Not for Private Gain organised opposition among governance scholars, former employees, and civil society groups.

On 5 May 2025 OpenAI changed course. The for-profit entity would become a Delaware public benefit corporation rather than a standard corporation, and the nonprofit would keep control.18 The restructuring completed on 28 October 2025, after both attorneys general signed memoranda of understanding with the company.19

The architecture of compromise

The resulting structure is a nested set of compromises, each answering a different constituency.

The Foundation holds 26 per cent of OpenAI Group PBC, the for-profit entity, a stake worth roughly $130 billion.20 Its governance powers are extensive on paper: it holds sole authority to appoint and remove every member of the corporation's board.21 The Safety and Security Committee, which can "require mitigation measures—up to and including halting the release of models or AI systems—even where the applicable risk thresholds would otherwise permit release", sits under the Foundation rather than the corporation.22

Microsoft holds 27 per cent, worth roughly $135 billion.23 It keeps exclusive distribution rights through Azure and remains the primary cloud partner. The terms shifted, though: OpenAI is no longer bound by a Microsoft right of first refusal on computing capacity, and has contracted to buy $250 billion in Azure services.24 Microsoft holds no governance power at all, with no board seats, no veto, and no formal role in safety or mission oversight.

Employees and other investors hold the remaining 47 per cent.25

The attorneys general extracted twenty concessions before approving the restructuring.26 Both entities must stay headquartered in California. The corporation's stated mission must be identical to the Foundation's. The Safety and Security Committee must be able to halt product releases. Both states promised continued monitoring.27

The structure looks robust on paper. Whether it is robust in practice turns on questions corporate law cannot answer.

What a public benefit corporation is, and is not

A Delaware public benefit corporation is a relatively recent legal invention. Under Subchapter XV of the state's General Corporation Law, such a company must name a specific public benefit in its charter, and its directors must balance three things: the financial interests of stockholders, the interests of those materially affected by the company's conduct, and the stated public benefit.

That is a real departure from the standard corporate form, where directors' duties run primarily to stockholders and are judged by shareholder value. In a public benefit corporation, directors who sacrifice some shareholder return in favour of the stated benefit are, in principle, protected from the lawsuits that would otherwise follow.28

The protection is real and untested. As of late 2025 no shareholder had sued over how a public benefit corporation managed its competing obligations.29 There is no case law on how courts should weigh financial interest against public benefit when the two conflict. The balancing standard, that a director's decision must be "informed and disinterested and not such that no person of ordinary, sound judgment would approve", grants wide discretion.30 Discretion is an asset when directors act in good faith and a vulnerability when they do not.

Several notable companies have adopted the form. Patagonia's 2022 conversion, in which Yvon Chouinard transferred ownership to a purpose trust and a nonprofit, gave the model credibility as a tool for protecting a mission at scale.31 Kickstarter reincorporated in 2015. Lemonade went public as one in 2020. None approached OpenAI's valuation, none operated in an industry of comparable geopolitical weight, and none faced the particular tension between safety and commercial pressure that defines frontier AI.

OpenAI Group PBC is by a wide margin the most valuable public benefit corporation ever created. It is a governance experiment without precedent and without a fallback.

The overlap problem

The structural criticism is not that the wrong form was chosen. It is that the form's safeguards are undermined by the people occupying its positions.

Nearly all board members serve on both the Foundation and the corporation.32 The Foundation is meant to oversee the corporation. The corporation is meant to be overseen by the Foundation. When the same individuals sit on both sides, oversight collapses into self-supervision.

Alnoor Ebrahim, who studies nonprofit accountability and the governance of social enterprises at Tufts, treats this as a fundamental governance failure.33 The Foundation holds 26 per cent equity, substantial in dollars and only a quarter of the entity it nominally controls. Ebrahim sets that against the Health Net precedent, where 80 per cent of equity went to a health foundation during a conversion of this kind, and against The Philadelphia Inquirer, where a nonprofit kept majority control.34

Judith Bell, an impact officer at the San Francisco Foundation, put it more directly: "There's a bazillion conflicts of interest here."

Steven Adler, a former leader on OpenAI's safety team, doubted the Safety and Security Committee's practical independence. The committee holds formal authority to halt model releases, and formal authority and operational independence are different things. "I hope that a truly independent body will do a better job of protecting the organization's mission than one that feels any pull toward profits," Adler said.35

The Eyes On OpenAI coalition, led by Orson Aguilar, was blunter: "The nonprofit continues to operate under the influence of the for-profit it supposedly oversees."36

The mission statement

OpenAI has revised its mission statement six times in nine years. Reviewing the tax disclosure covering 2024, released in November 2025, Ebrahim noticed that the word "safely" had gone.37

The earlier formulation read: "to build general-purpose artificial intelligence that safely benefits humanity, unconstrained by a need to generate financial return". The current version reads: "to ensure that artificial general intelligence benefits all of humanity". Both the safety commitment and the disavowal of financial motive have been dropped.38

The word "safely" was the anchor connecting the mission to the specific concern, the risk posed by advanced AI systems, that justified the nonprofit structure in the first place. Without it, the governing documents of both entities describe a mission that could be satisfied, on the language alone, by beneficial AI that is not safe, or by safe AI whose benefits accrue to investors rather than to humanity broadly. Ebrahim reads the change as a shift in priorities from safety toward commercial objectives, and holds that the attorneys general of California and Delaware allowed it to happen.39

OpenAI's leadership has not publicly explained the change. Its blog posts and structural documents from 2025 use the shorter formulation without comment, and the Safety and Security Committee's formal powers remain in place.

Words in mission statements are not decorative. They are the standard against which fiduciary duties are measured, board decisions justified, and regulatory oversight exercised. When the standard changes, the accountability framework changes with it.

The safety committee question

The Safety and Security Committee is the element most often cited as evidence that the restructuring preserved meaningful safety governance. Its formal powers are significant. It sits under the Foundation rather than the corporation, it is chaired by Zico Kolter, a machine learning researcher at Carnegie Mellon, and its mandate includes requiring mitigation up to and including halting a model release, even where the applicable risk thresholds would otherwise permit one.40

On paper this is the strongest safety governance mechanism at any major AI company. No other frontier laboratory has a committee empowered to halt a release unilaterally, housed in a separate legal entity from the commercial operation, and chaired by an independent researcher.

Whether formal power translates into operational reality is the open question, and it depends on three conditions. The committee needs timely access to the technical information required to assess risk, during development rather than once a model is ready to ship. Its members need the willingness to use that authority against the commercial interests of a $500 billion company whose employees hold 47 per cent of the equity. And the Foundation board needs the willingness to defend the committee's decisions when they are contested.

The corporate structure guarantees none of these. They rest on norms, culture, and individual resolve, which are precisely the factors that failed in November 2023.

The board that fired Altman had formal authority. It used that authority. Within days it was replaced by a board that would not. The Safety and Security Committee holds authority exceeding what that board had. Whether it would survive using it against the combined will of management, investors, and a workforce holding billions in equity remains untested.

Governance by design, tested by reality

Four months after the restructuring completed, the new structure met its first test.

On 27 February 2026 the United States government designated Anthropic, OpenAI's principal competitor in frontier AI, a supply chain risk and terminated its military contracts, after the company refused to drop safety restrictions covering autonomous weapons and mass surveillance. President Trump separately directed all federal agencies to stop using Anthropic's tools.41 Within hours, OpenAI announced it had secured the Pentagon contract.42

Altman moved quickly to manage the reaction. In a memo to staff he wrote that OpenAI would hold the same lines Anthropic had defended: no AI for mass surveillance of American citizens, and no fully autonomous lethal weapons.43 The contract was amended to state that "the AI system shall not be intentionally used for domestic surveillance of U.S. persons and nationals", which Altman framed as clarifying obligations the Fourth Amendment and existing law already imposed.44

By 3 March, Altman acknowledged the damage. "We were genuinely trying to de-escalate things and avoid a much worse outcome," he wrote, "but I think it just looked opportunistic and sloppy." He added that he should not have rushed to announce the deal on the same day Anthropic was sanctioned.

The episode exposes the limits of the structure. Altman's commitments are substantively close to the contractual terms that cost Anthropic its contract. Anthropic's were binding provisions, enforceable under the agreement. Altman's are public commitments, voluntarily adopted and voluntarily amendable. The amended language covers domestic surveillance of US persons, where Anthropic's restrictions were broader, covering mass surveillance and autonomous weapons targeting without geographic limit.45

The corporate structure did not prevent OpenAI from taking the contract. The Safety and Security Committee did not publicly weigh in. The Foundation board did not intervene. Twenty concessions from two attorneys general, a mission-aligned charter, and an empowered safety committee were not visibly activated.

That is not necessarily a failure. The committee may have reviewed the contract and found it acceptable; the Foundation board may have considered the matter and concluded that a Pentagon contract carrying safety restrictions served the public benefit. The absence of any visible governance process is itself informative, though. When the first significant test arrived, the public response came from the chief executive's social media account rather than from the mechanisms designed for exactly that purpose.

The condition not yet tested

The OpenAI restructuring is the most ambitious attempt yet to build mission-driven governance into a commercial AI company. It is more sophisticated than any previous conversion of this kind, and the Foundation's formal powers over board appointment, safety authority, and mission alignment exceed what any other AI company's structure provides.

Formal powers are scaffolding rather than architecture. The November 2023 crisis showed that a board with authority to remove a chief executive could itself be removed for using it. The February 2026 episode showed that a safety committee with authority to halt releases did not visibly engage when the company's most consequential policy decision was taken within hours.

The critics are right about the overlap problem, the mission statement, and the enforcement gaps. The defenders are right that the Foundation's powers are real, that the committee's mandate is the strongest in the industry, and that two attorneys general are actively monitoring compliance. Both can be true at once. The structure is the best available, and it may also be insufficient.

The question the restructuring poses is not whether a public benefit corporation is better than a conventional one for governing frontier AI. It plainly is. The question is whether any corporate form, however carefully designed, however many concessions regulators extract, however empowered its safety mechanisms, can hold the tension between a $500 billion valuation and a mission that may require saying no to the most powerful customer in the world.

The November 2023 board had both the power and the conviction. It lasted five days. The Safety and Security Committee has the power. Whether it will ever need the conviction, and what happens if it does, will determine whether this experiment is remembered as a governance success or as the most expensive illustration of mission drift in corporate history.

Footnotes

  1. OpenAI, "Introducing OpenAI," 11 December 2015. ↩

  2. Additional co-founders included Trevor Blackwell, Vicki Cheung, Andrej Karpathy, Durk Kingma, John Schulman, Pamela Vagata, and Wojciech Zaremba. ↩

  3. Donors pledging the initial $1 billion included Altman, Musk, Greg Brockman, Reid Hoffman, Jessica Livingston, Peter Thiel, Amazon Web Services, and Infosys. ↩

  4. OpenAI, "OpenAI LP," 11 March 2019. ↩

  5. The hundredfold cap applied to initial investors in the capped-profit entity. Later rounds negotiated different terms. ↩

  6. Microsoft, "Microsoft invests in and partners with OpenAI to support us building beneficial AGI," 22 July 2019. ↩

  7. CNBC, "Microsoft to invest $10 billion in ChatGPT creator OpenAI, report says," 10 January 2023. ↩

  8. Microsoft's cumulative investment of approximately $13.8 billion covered the 2019 investment, the 2023 commitment, and interim rounds. ↩

  9. Helen Toner was then a researcher at Georgetown's Center for Security and Emerging Technology. ↩

  10. OpenAI, board statement, 17 November 2023. ↩

  11. Axios, "OpenAI chaos: A timeline of Sam Altman's firing and return," 22 November 2023. ↩

  12. ABC News, "4 days from fired to re-hired: A timeline of Sam Altman's ouster from OpenAI," 22 November 2023. ↩

  13. Ilya Sutskever, post on X, 20 November 2023. ↩

  14. PBS News, "Sam Altman reinstated as OpenAI CEO with new board replacing the one which fired him," 22 November 2023. ↩

  15. Sutskever announced his departure on 14 May 2024 and subsequently co-founded Safe Superintelligence Inc. ↩

  16. ProMarket, "OpenAI Abandons Move to For-Profit Status After Backlash. Now What?" 6 May 2025. ↩

  17. The California attorney general's scrutiny rested on the state's authority over charitable assets under the Supervision of Trustees and Fundraisers for Charitable Purposes Act. ↩

  18. OpenAI, "Why OpenAI's structure must evolve to advance our mission," 5 May 2025. ↩

  19. The California memorandum was signed 27 October 2025; Delaware issued a statement of no objection on 28 October 2025. ↩

  20. EA Forum, "The OpenAI Governance Transition: The History, What It Is, and What It Means," October 2025. ↩

  21. The Foundation's sole authority to appoint and remove board members was confirmed in the memoranda of understanding with both attorneys general. ↩

  22. OpenAI, "Our structure," accessed October 2025. ↩

  23. Microsoft, "The next chapter of the Microsoft–OpenAI partnership," 28 October 2025. The stake is stated on an as-converted, diluted basis. ↩

  24. Ibid. ↩

  25. EA Forum, "The OpenAI Governance Transition," October 2025. ↩

  26. CalMatters, "OpenAI's restructuring deal with California is full of holes, critics say," October 2025. ↩

  27. Both attorneys general committed to ongoing monitoring as part of their non-objection. ↩

  28. Delaware General Corporation Law, Title 8, Chapter 1, Subchapter XV, §365. ↩

  29. As noted by the American Bar Association and Delaware corporate law practitioners, no shareholder suit had tested the balancing standard as of late 2025. ↩

  30. Delaware Code, Title 8, §365. ↩

  31. Patagonia transferred all voting stock to the Patagonia Purpose Trust and all nonvoting stock to the Holdfast Collective in September 2022. ↩

  32. Alnoor Ebrahim, The Conversation, 2026. ↩

  33. Ebrahim is Professor of International Business at the Fletcher School and Tisch College of Civic Life, Tufts University. ↩

  34. Ebrahim cites Health Net, where 80 per cent of equity went to a health foundation, and The Philadelphia Inquirer, where a nonprofit retained majority control. ↩

  35. Steven Adler, quoted in CalMatters, October 2025. ↩

  36. Orson Aguilar, director of the Eyes On OpenAI coalition, quoted in CalMatters, October 2025. ↩

  37. Ebrahim identified the change while reviewing OpenAI's IRS disclosure covering 2024, released in November 2025. ↩

  38. The current statement appears on openai.com and in the corporation's charter documents. ↩

  39. Ebrahim, The Conversation, 2026. ↩

  40. OpenAI, "Our structure," accessed October 2025. Kolter also serves on the Foundation board. ↩

  41. Supply chain risk designation, 27 February 2026, alongside a directive that federal agencies cease using Anthropic's tools. ↩

  42. CNBC, "OpenAI strikes deal with Pentagon, hours after rival Anthropic was blacklisted by Trump," 27 February 2026. ↩

  43. Axios, "Sam Altman says OpenAI shares Anthropic's red lines in Pentagon fight," 27 February 2026. ↩

  44. CNBC, 3 March 2026. ↩

  45. Anthropic's acceptable use policy prohibited mass surveillance and fully autonomous weapons targeting without restriction to US persons or domestic contexts. NPR, "OpenAI says it shares Anthropic's 'red lines' over military AI use," 27 February 2026. ↩