This article is part of our series that explores the business of artificial intelligence
On Monday, OpenAI CEO Sam Altman announced that his company would soon be releasing their first open language model since GPT‑2 in the coming months. “We’ve been thinking about this for a long time but other priorities took precedence. now it feels important to do,” Altman said.
This is happening while OpenAI has become virtually the opposite of its name (with the exception of Whisper, its speech-to-text model) since the release of GPT-3. There could be multiple reasons for the change of heart in OpenAI and its leadership. Here are three potential reasons and what it means for the future of the AI industry.
Pressure from the open-source community
For the most part of the past few years, open models have been on the heels of state-of-the-art closed models, usually one generation behind. But things changed with the release of DeepSeek-R1, an open-weight and permissively licensed reasoning model that was nothing short of OpenAI’s o1.
In addition to matching the quality of o1, DeepSeek-R1 undercut OpenAI’s secrecy by fully revealing its reasoning chain, making it a much more useful model. Soon after its release, it was hosted on many major cloud and neo-cloud providers, and began appearing on mass consumer products such as Perplexity, which previously used o1 for its advanced search capabilities.
Since its release, R1 has been derived and adapted thousands of times by different developers. I’m seeing more and more research papers using R1 as the base model to fine-tune for their experiments. Basically, R1 proved that open source is on par with closed models and is now setting the tone for research going forward.
Soon after the release of R1, Altman admitted in a Reddit AMA, “I personally think we have been on the wrong side of history here and need to figure out a different open source strategy.”
The industry is maturing
As I mentioned in a recent article on AI moats, OpenAI has the advantage of high brand quality and being the first mover in the field. For many people, ChatGPT is their first experience of modern AI and their first go-to platform to learn how to use large language models (LLMs). ChatGPT has the biggest user base in the consumer AI market.
However, as people move from exploration to adoption, many will want to integrate LLMs into their applications. These enterprise use cases present a bigger business opportunity, where companies pay on a per-token or GPU usage basis.
The enterprise sector requires more flexibility on the part of the model providers. Some will be fine with connecting their applications to third-party services such as OpenAI’s API. Others will require hosting the model on their own cloud or on-prem servers. And some will start with a third-party API but will want to migrate to their own GPUs to scale and reduce costs. OpenAI currently can’t serve the second and third types of customers, considerably narrowing its total addressable market.
Having an open source model will give OpenAI more flexibility. Customers can start with the closed models and gradually move to compatible open models as their applications mature.
OpenAI might be going solo?
Until now, OpenAI has been mostly relying on Microsoft to host its models. But there is a growing rift between the two companies and OpenAI is currently shopping for other cloud providers. But Altman is very ambitious and is considering creating its own AI accelerators and potentially data centers. It might eventually host and serve its own models.
If such a thing happens, OpenAI will be able to copy Google’s playbook, providing a wide range of options, from its closed models (GPT-4o and o1) to its upcoming open source models. This will be convenient for developers, who will be able to explore all kinds of configurations.
While this is very speculative for the moment, having and open source strategy will provide OpenAI with stability in the uncertain times it faces in the upcoming months and years, whether it is from its partners or the sudden shock of new open source models.