‘Selling coffee beans to Starbucks’ – how the AI boom could leave AI’s biggest companies behind

The importance of foundation models in AI has become a topic of intense discussion among startups, many of which are now comfortable with what were once considered mere 'GPT wrappers.' These companies are increasingly focused on customizing AI models for specific applications and developing user interfaces, viewing the foundational models as interchangeable commodities. This trend was highlighted at the recent Boxworks conference, where the spotlight was on software applications built on top of AI models. A key factor driving this shift is the diminishing returns from the pre-training phase of AI development—the initial stage where models learn from vast datasets, a process managed solely by foundation model companies. While AI continues to advance, the initial advantages of large-scale foundational models are waning, leading to a pivot towards post-training techniques and reinforcement learning as the next frontiers. For instance, if the goal is to enhance an AI coding tool, developers might find greater success in refining the model’s performance and interface design rather than investing billions in additional pre-training. The emergence of successes like Anthropic’s Claude Code illustrates that companies focusing on these areas can thrive without solely relying on foundational model development. As the AI competitive landscape evolves, the dominance of major labs such as OpenAI and Anthropic is increasingly being challenged. The focus is shifting away from a singular quest for an all-encompassing AGI; instead, we are witnessing a proliferation of specialized businesses in various domains like software development and data management. The once-coveted first-mover advantage is now under scrutiny, with startups finding it feasible to transition between different foundational models seamlessly, diminishing the significance of any single model’s superiority. In this transformed environment, the abundance of open-source models could render foundation model companies vulnerable, potentially relegating them to the role of backend suppliers in a low-margin market. As one industry founder put it, this could mean these firms end up akin to "selling coffee beans to Starbucks." The implications of this shift are substantial. The AI boom has historically been tied to the success of companies developing foundational models, with giants like OpenAI and Google at the forefront, leading many to believe their influence would be long-lasting. However, as we have seen over the past year, the emergence of robust third-party AI services is changing the game. These services often utilize various foundational models interchangeably, making it irrelevant for startups which model they choose. Despite the challenges, foundation model companies still wield significant advantages, including brand recognition and substantial financial resources. While OpenAI's consumer products may be more challenging to replicate than its coding models, the landscape remains dynamic, and the interest in post-training approaches might evolve rapidly. As we stand on the brink of new breakthroughs in AI, particularly in fields like pharmaceuticals and materials science, the future remains uncertain. The trend toward building larger foundation models appears less attractive than before, highlighting the risks involved in the ongoing AI arms race.