Argues the system that almost broke wasn't the technology but the assumption that a frontier AI lab could be governed like a small academic nonprofit. Points to the absence of a written rationale, no transition plan, no successor briefed, and no communication to Microsoft as evidence that the governance model was structurally unfit for infrastructure of this scale.
Brockman repeatedly declines to relitigate motives and says the directors 'believed they were doing the right thing,' but his operational account is damning: the 'not consistently candid' phrase was workshopped after the fact, there was no successor briefed, and the board apparently expected employees to shrug and keep shipping. He frames the failure as one of execution and assumptions about institutional inertia, not malice.
Reframes the story away from boardroom drama and toward infrastructure risk, noting the 72-hour window was the closest a piece of generally-available developer infrastructure has come to disappearing overnight since left-pad. The lesson for developers is structural: production traffic was running through OpenAI endpoints with no contingency for the provider ceasing to exist.
By surfacing the Brockman interview to the HN front page (189 points, 194 comments), signals that the developer community views this as a technical-infrastructure retrospective worth dissecting, not just celebrity executive drama. The submission framing emphasizes the '72 hours that almost killed OpenAI' angle.
Greg Brockman sat down with Shane Parrish on The Knowledge Project for what is, two and a half years later, still the most candid first-person account of the weekend OpenAI nearly imploded. On November 17, 2023, the OpenAI nonprofit board — Helen Toner, Tasha McCauley, Adam D'Angelo, and Ilya Sutskever — fired Sam Altman over a Google Meet, then removed Brockman as chairman in a second call minutes later. Brockman resigned within hours. By Monday morning, 738 of roughly 770 employees had signed an open letter threatening to follow Altman and Brockman to Microsoft, where Satya Nadella had quietly offered to absorb the entire company.
Five days later, Altman was reinstated, the board was reconstituted with Bret Taylor and Larry Summers, and Sutskever publicly recanted. Brockman's telling adds texture the contemporaneous reporting didn't have: the board never produced a written rationale, the "not consistently candid" phrase was workshopped after the fact, and the directors apparently expected employees to shrug and keep shipping. The system that almost broke wasn't the technology — it was the assumption that a frontier AI lab could be governed like a small academic nonprofit.
Brockman is careful not to relitigate motives. He repeatedly says he doesn't know exactly what triggered the firing and that the board members "believed they were doing the right thing." But the operational story he tells is brutal: no transition plan, no successor briefed, no communication to Microsoft (which owned 49% of the for-profit and learned via news alert), no plan for the API customers running production traffic through OpenAI's endpoints at the moment the announcement hit.
It's tempting to file this under "juicy boardroom drama" and move on. Don't. The 72-hour window is the closest a piece of generally-available developer infrastructure has come to disappearing overnight since the npm left-pad incident, and the lessons are structural, not gossip.
First, the governance model. OpenAI's nonprofit board had four independent directors with no equity, no operating role, and — by Brockman's account — no real workflow for evaluating the CEO of a company shipping models that 100 million people used weekly. Compare that to the boards governing AWS, Azure, or GCP infrastructure: heavily regulated public companies with fiduciary duties to shareholders, audit committees, and decades of crisis playbooks. We've been treating Anthropic, OpenAI, xAI, and Mistral as utilities while their governance still looks like a Series A startup's. That gap is closing, but slowly.
Second, the resilience mechanism that actually worked was the employee letter, not the board structure. Brockman is explicit that the org survived because the staff collectively had a credible exit option — Microsoft would have hired them all, the IP would have been re-implemented, and the nonprofit would have inherited an empty office. That's not a repeatable governance solution. It's a coincidence of having a major-cloud parent with infinite headcount and a CTO willing to move on a Friday night. The next lab in crisis won't have a Nadella waiting.
Third, the customer side of the equation was barely a factor in the board's thinking. Brockman recounts triaging API uptime concerns while simultaneously trying to figure out if he still worked at the company he co-founded. Enterprise customers — including Microsoft's own Copilot pipeline — had no SLA carve-out for "founder fired, company dissolving." The post-crisis result is the wave of multi-provider abstraction layers we've seen since: LiteLLM, OpenRouter, Bedrock's cross-model routing, the explosion of "model-agnostic" middleware. The November weekend was the inflection point where "never single-vendor your LLM" stopped being paranoia and became table stakes.
Fourth, and this is the part most coverage missed: Brockman describes a deep cultural fracture between the safety-oriented research staff and the product/scaling org that the crisis surfaced but didn't resolve. Sutskever left within months. Jan Leike left. The superalignment team was disbanded. The board crisis was an early symptom of a question the field still hasn't answered — who's allowed to slow down a frontier lab, and what does the override mechanism look like? In OpenAI's case the answer turned out to be "nobody, and there isn't one."
If you're shipping anything that depends on a frontier model API, the November 2023 weekend should already be in your risk register. Concrete actions, in rough order of leverage:
Abstract your provider. If you're calling `openai.chat.completions.create` directly from app code in 2026, you're carrying technical debt that has a specific dollar value attached to it. Wrap calls behind an internal interface — even a thin one — so a provider swap is a config change, not a refactor. Tools like LiteLLM, Portkey, and AWS Bedrock have eaten this complexity for you; there's no longer a build-vs-buy argument here.
Keep a warm fallback model. "Warm" means you've actually tested your prompts against Claude, Gemini, or an open-weights model running on Groq or Together. Prompt portability is real but imperfect; you'll discover the rough edges only by running your evals against the alternate stack at least quarterly. The teams who came out of the OpenAI weekend looking smart were the ones with a 30-minute failover, not the ones with a 30-page incident plan.
Read the governance docs of the labs you depend on. Anthropic has the Long-Term Benefit Trust. OpenAI has its post-2023 reconstituted board and an in-progress conversion to a public benefit corporation. xAI is a Musk-controlled entity. These structures determine whether your vendor can be fired by five people on a Friday afternoon. Treat board composition as an SLA component.
Negotiate continuity clauses. Enterprise contracts post-2023 increasingly include change-of-control and continuity-of-service language. If yours don't, ask. Even a clause requiring 30 days' notice of "material governance changes" gives your incident team time to flip the failover flag.
The Brockman interview lands at a useful moment. OpenAI is mid-conversion to a PBC, Anthropic just raised at a number that makes Microsoft's 2023 stake look quaint, and the EU AI Act's governance provisions for general-purpose models kicked in last year. The next twelve months will tell us whether the labs have actually professionalized their governance or just papered over it with new logos on the board page. Brockman's account is a reminder that the answer matters to anyone with a production system on the other end of an API key — which, in 2026, is most of us.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.