In November 2023, OpenAI’s board fired Sam Altman on a Friday. By Monday he was back. 700 employees had threatened to quit. Microsoft had intervened. The safety-focused non-profit board, designed to govern the world’s most powerful AI, had lasted 96 hours against commercial reality.
July 2026 · Module 2 · Course Case Study
As you read — hold this question
Can a governance structure designed for a research lab survive when the lab becomes a $29 billion commercial entity with 100 million users?
96 hrs
from fired to reinstated. The non-profit safety board tried to govern the world’s most commercially valuable AI company. It lost in four days.
OpenAI was founded in 2015 as a non-profit with a mission to develop AI “for the benefit of humanity.” By 2023, it had a capped-profit subsidiary valued at $29 billion, a $13 billion investment from Microsoft, and 100 million ChatGPT users. The board that fired Altman cited “loss of candour.” Altman’s supporters cited loss of commercial momentum. The episode revealed something the Module 2 technical content implies but doesn’t state: who governs AI is as consequential as how it is built.
The technical stack — connecting to cases 1–5
Quiz: All Module 2
What OpenAI actually built — and which principles from this module it illustrates
OpenAI product / technique
Module 2 principle
Case reference
GPT pre-training on internet text — predict the next token
Self-supervised learning at scale; emergence from scale
Case 1 (learning paradigms), Case 3 (transformers)
RLHF — human raters score responses, reward model trained, RL fine-tunes
Reinforcement learning from human feedback as alignment layer
Case 1 (RLHF), Case 5 (alignment)
GPT-4 as foundation model underlying ChatGPT, Copilot, and 3rd party APIs
Homogenisation — one model, thousands of downstream applications
Case 4 (Storey et al., homogenisation)
Emergent capabilities — GPT-4 passes bar exam without legal training
Emergence — capabilities not explicitly trained appear at scale
Prediction ≠ comprehension; model cannot verify its own outputs
Case 3 (language barrier)
The board crisis — November 2023
96 hours that redefined AI governance
Fri 17 Nov 2023
OpenAI board fires Sam Altman, citing “lack of candour.” President Greg Brockman resigns. No public explanation of the safety concern.
Sun 19 Nov (morning)
Microsoft CEO Satya Nadella announces Altman and Brockman will join Microsoft to lead a new AI research lab. OpenAI’s commercial future in doubt.
Sun 19 Nov (evening)
700+ of OpenAI’s 770 employees sign letter threatening to quit and join Microsoft unless the board resigns and Altman is reinstated.
Mon 20 Nov 2023
Sam Altman reinstated as CEO. Board reconstituted with new members. The safety-focused directors who voted to fire him are gone.
2024–25
OpenAI raises $6.6B at $157B valuation in 2024. Converts to for-profit public benefit corporation in 2025. Non-profit retains minority stake.
What the crisis revealed
The OpenAI board was designed to prioritise safety over commercial interests. It had the legal authority to fire the CEO. It exercised that authority — and was overridden within 96 hours by commercial pressure from employees, investors, and Microsoft. The non-profit structure did not fail because of poor design. It failed because at $29B valuation with 100 million users, the commercial stakes exceeded the governance structure’s ability to resist them.
The course case study — strategic analysis
OpenAI’s commercial AI strategy: from research lab to platform
The BUSN9049 case study examines how OpenAI moved from an academic research organisation to the world’s most commercially significant AI company in less than 10 years. The technical capabilities (cases 1–5) are the input. The commercial strategy is what transformed them into a product.
The platform strategy
GPT API: one foundation model, 3M+ developer applications built on top
ChatGPT as consumer product: 100M users in 2 months — fastest product adoption in history
Microsoft partnership: $13B and integration into Office, Azure, GitHub, Bing
Homogenisation in practice: OpenAI’s model underlies most enterprise AI deployments
The governance tensions
Non-profit mission vs capped-profit commercial entity: competing incentives by design
Safety as a product differentiator vs safety as a constraint on commercialisation
Homogenisation risk: OpenAI’s biases and failure modes propagate to every downstream application
Who governs the model that governs the most deployed AI products?
The Principle — technical capability creates governance requirements
The OpenAI case connects every Module 2 concept to organisational and commercial reality. Foundation models (Case 4) create homogenisation risk — whoever controls the foundation model controls the values, biases, and failure modes of thousands of downstream applications. RLHF alignment (Case 1) reflects the preferences of the human raters OpenAI selects — not a neutral baseline. The EU AI Act (Case 5) imposes accountability on deployers, but foundation model developers remain less regulated. The board crisis is not a leadership story. It is a governance story about who should have authority over systems that operate at civilisational scale.
Take this away
The OpenAI board had the legal authority to govern the world’s most powerful AI. It exercised that authority — and was overridden in 96 hours by commercial pressure. Governance structures for transformative AI must be designed before the commercial stakes make good governance impossible to enforce.
Trace the November 2023 board crisis timeline — and identify the moment the safety mission lost.
Fri 17 Nov: Board fires Altman citing “lack of candour” — no public explanation of the safety concern. Brockman resigns. Sun 19 Nov morning: Nadella announces Altman will join Microsoft. Sun 19 Nov evening: 700+ of OpenAI’s 770 employees sign letter threatening to quit. Mon 20 Nov: Altman reinstated; safety-focused board members gone. The mission lost Sunday night: when 700 employees threatened to follow Altman to Microsoft, the commercial stakes exceeded the governance structure’s ability to resist. The board had legal authority. It could not survive economic leverage.
Question 2 of 3
Identify which Module 2 concept is illustrated by each of OpenAI’s key technical or commercial decisions.
GPT pre-training → self-supervised learning at scale + emergence (Cases 1, 3). RLHF → reinforcement learning from human feedback as alignment layer (Cases 1, 5). GPT-4 as foundation model underlying ChatGPT, Copilot, and 3rd party APIs → homogenisation (Case 4). Emergent capabilities: GPT-4 passes bar exam without legal training → Storey et al. emergence (Case 4). ChatGPT hallucinations → prediction ≠ comprehension (Case 3).
Question 3 of 3
The non-profit board “did not fail because of poor design — it failed because at $29B valuation with 100M users, the commercial stakes exceeded the governance structure’s ability to resist them.” What does this tell you about when governance structures need to be designed?
Governance structures need to be designed before the stakes make good governance impossible to enforce — not after the organisation reaches the scale at which commercial pressure can override safety-focused authority. OpenAI’s board had the right structure (mission-driven, fiduciary to humanity) but was designed for a research lab, not a $29B commercial entity. By the time the board tried to exercise its authority, the economic leverage of employees, investors, and Microsoft had already outweighed its legal power. The implication: governance of transformative AI must be built into the architecture of the organisation before, not after, it becomes commercially transformative.
Sources
OpenAI history
Altman, S. et al. (2015). OpenAI founding blog post. openai.com.
Board crisis
Metz, C. & Weise, K. (2023, November). Inside the drama that upended OpenAI. New York Times.
Employee letter
OpenAI staff letter (2023, November 19). Open letter to the OpenAI board. Reported by The Verge.