The OpenAI board crisis

The board believed it was playing a one-shot principal-agent game. Hire the CEO, fire the CEO, replace the CEO. Standard board prerogative under the OpenAI charter. The legal authority was uncontested.

What was being played was a coalition-formation game with three implicit veto players the board had not budgeted for: Microsoft, the employees, and the tender offer that gave the employees a near-term financial reason to act collectively.

The information asymmetry ran in the opposite direction from what the board assumed. The board held private information about Sam (the alleged candour failures). Microsoft and the employees held private information about the cost of his removal: Nadella had a Saturday morning offer ready, the employees had a Slack channel and a draft letter. None of that was visible from the boardroom on Friday afternoon.

The game was sequential. Friday: the board moves. Saturday: Microsoft moves. Sunday: the employees move. Monday: the board faces a coalition it did not see forming. By Tuesday, the equilibrium had flipped and the board was replaced.

Each player's payoff pointed in a different direction. None of them mapped cleanly onto the others, and the gap between stated and actual goals is where the board lost the game.

• The board: Stated goal: preserve the OpenAI mission and safety mandate. Actual goal: reassert control over the direction of the leading AI lab. BATNA: walk away as the directors who blocked the AI race. Career-ending in tech. Time horizon: long and mission-oriented. They underweighted the next 72 hours completely.
• Sam Altman: Stated goal: continue leading OpenAI through the GPT-4 era. Actual goal: lead whichever entity becomes the leading AI lab. The label matters less than the position. BATNA: a Microsoft AI lab with cash, compute, and any employee who follows. Strong. Time horizon: compressed to hours by the public nature of the firing.
• Microsoft (Satya Nadella): Stated goal: supportive of any outcome that protects the partnership. Actual goal: secure access to the talent and IP regardless of which legal shell holds them. BATNA: absorb the talent into a Microsoft AI lab. Slower, but safe. Time horizon: quarters. The race outlasts any single weekend.
• The 770 employees: Stated goal: continue the work, regardless of who runs the company. Actual goal: protect the upcoming tender offer that would make many of them millionaires. BATNA: Microsoft would hire any of them, at the same compensation, immediately. Time horizon: long career, but the tender offer was on the calendar.

The board's failure was not in values. It was in the model of the game. They did not assign a payoff to the response of Microsoft or the employees, because they did not see those parties as players. In a principal-agent game, the agent has no allies. In the actual game, the agent had two coalitions waiting in the wings.

By Sunday evening, the board's options had collapsed. Apply iterated dominance to what remained, modelling the best response from Sam, Microsoft, and the employees against each move.

Three strategies are strictly dominated once you model the response of Sam, Microsoft, and the employees. Only one move survives iterated elimination: accept reinstatement and negotiate the board's own replacement on terms rather than under duress.

Run the game forward with mission preservation as the real goal. This is the retrospective recommendation for the board, and it applies symmetrically to any board considering the removal of a charismatic founder.

• Primary move: Negotiate the exit privately first. A public firing creates a public counter-coalition. The board's authority was credible only as long as the move stayed inside the boardroom.
• Commitment device: Pre-secure a successor with employee credibility. Mira Murati was loyal to Sam. Naming her would have signalled continuity. Naming Emmett Shear signalled rupture.
• Signal: Frame the action as governance reform, with a published timeline. A clear story removes Microsoft's incentive to intervene as a saviour.
• Coalition pre-emption: Brief the top fifteen technical leads before announcing. Their public support, or even their public neutrality, breaks the employee coalition before it forms.
• Contingency: If Microsoft objects, have a board statement ready that frames the partnership as unaffected. Ambiguity is the friend of the firing party. Silence is fatal.
• Timing: Not on a Friday afternoon before the US Thanksgiving week. The empty calendar gives the counter-coalition 72 free hours. Move on a Wednesday morning, with a planned all-hands within four hours.

Key statistics from the actual outcome: 770 of 800 employees signed the letter, the equilibrium flipped in 96 hours, and 3 of 4 original board members lost their seats. The mission itself survived unchanged.

Two years on, the secondary effects are visible. The game did not end when Sam walked back in. It propagated through every other lab that watched the board lose.

Other AI labs studied the game and built different governance. Anthropic kept its Long-Term Benefit Trust but paired it with founder-supportive bylaws. xAI never had a non-profit veto. Mistral incorporated as a normal European company. The board veto became a known liability in AI lab structuring.

Microsoft demonstrated that its power extended past the contract. The lesson registered with every other lab dependent on hyperscaler compute. The next generation of AI labs raised more capital, faster, partly to escape the same structural dependency.

The board members, in the year after, became cautionary references rather than authorities. The reputation cost turned out to be the dominant cost of the entire episode: larger than any short-term governance victory could have justified. This reframes the payoff matrix retrospectively. Had the board correctly weighted reputation as an asset in their own payoff function, the dominant strategy changes before Friday even arrives.