I was surprised by how quickly the OpenAI drama unfolded. Sam Altman got fired on November 17th and was back as CEO five days later. The board that fired him mostly resigned, and the whole thing happened so fast it made me think about the risks of depending on OpenAI at enterprise scale.
The OpenAI board fired Altman, citing a lack of candour, but that decision backfired. Almost all of OpenAI's staff signed a letter threatening to resign if Altman was not reinstated, and Microsoft, which had invested $13 billion, stated it would hire Altman and the departing team. The board had no choice but to capitulate, and a new board was formed with Altman returning as CEO.
OpenAI's governance structure is complex. It operates as a non-profit that controls a capped-profit entity, and the board of the non-profit has a fiduciary duty to the mission of beneficial AI, not to investors or revenue. This structure is why Microsoft's $13 billion investment did not give them board control, but it also means the board can act against commercial interests if they believe safety requires it.
For instance, I have seen this play out in production environments where companies like Amazon with SageMaker or Google with Vertex AI provide a more traditional cloud-based service level agreement, which can be more appealing to enterprises that require high levels of reliability and stability. In contrast, OpenAI's unique governance structure can be a double-edged sword, providing a high level of autonomy but also introducing uncertainty that can be challenging to navigate.
The tension between commercial pressure and safety-focused governance is structural, and it's not resolved by Altman's return. This tension is a fundamental aspect of OpenAI's governance, and it's something that enterprises need to consider when deciding how much to depend on OpenAI's APIs. To mitigate this risk, companies can use tools like TensorFlow or PyTorch to develop their own models, which can provide more control over the development and deployment process, but at a higher cost and resource requirement.
The implications of this governance instability are significant for enterprises with significant revenue or operational dependency on OpenAI APIs. The fact that OpenAI can experience governance instability significant enough to threaten the company's continuity within a five-day window is a data point that enterprises cannot ignore. In my experience, this kind of instability can result in downtime costs of up to $1 million per hour for large enterprises, making it crucial to have a contingency plan in place.
Using a service like Azure OpenAI Service can help mitigate some of this risk by providing an additional layer of abstraction and support. For example, Azure provides a 99.99% uptime guarantee, which can help reduce the risk of downtime and provide more stability for enterprises. However, this comes at an additional cost, and companies need to weigh the trade-offs between cost, reliability, and control when deciding how to proceed.
So what can enterprises do to mitigate this risk? Having a model diversity strategy is key. This means developing applications that can route to alternative models if needed, and evaluating them regularly, not just planning for it. It's not about abandoning OpenAI APIs, but about being prepared for any eventuality.
The OpenAI crisis also had an interesting consequence for Microsoft. The company came out of the crisis with more effective influence, not less. Altman is back as CEO with backing from Microsoft, and the new board has investor representation. This means that Azure OpenAI Service, which routes enterprise OpenAI usage through Microsoft's infrastructure, becomes more attractive as a compliance and continuity layer between enterprises and OpenAI's governance uncertainty.
In the end, the OpenAI drama highlights the complexities of governance in AI companies. It's a reminder that even the most successful companies can experience significant instability, and that enterprises need to be prepared for any eventuality. By understanding the governance structure of OpenAI and the implications of its instability, enterprises can make more informed decisions about their dependency on OpenAI's APIs.