Large Language Models: Fragmented Market or the Winner Takes it All? : Trustworthy Emerging Technologies, Winter Term 23/24

Background: Since late 2022, Large Language Models (LLMs) from major players like OpenAI, Anthropic, Google, and Meta have advanced significantly, prompting reflection on AI's societal impact. Concerns about monopolistic trends in tech, exemplified by companies like Microsoft, Alphabet, and Meta underscore the need to scrutinize market dynamics. The integration of Generative AI (GenAI) into workflows raises questions about consolidation, with OpenAI's platformization efforts indicating a potential trend toward monopolization.
Objective: This research paper aims to analyze the market structure of the GenAI landscape, focusing on Large Language Model Foundation Providers (LLMFPs) and Large Language Model Layer Providers (LLMLPs). Despite recognizing risks, research on consolidation trends in the LLM market is limited, motivating this study to analyze current dynamics and assess potential monopolistic or oligopolistic outcomes.
Methods: The methodology employed in this research paper involves a qualitative approach utilizing interviews with industry experts. The final sample size consisted of eight interviewees who are active as investors, consultants or entrepreneurs in the field of GenAI. ... mehrData analysis revolved around a robust coding framework, with selective coding as the primary approach. Seven key concepts of market structure served as initial codes, and 156 text passages from the transcripts were assigned to these codes.
Results: The research findings suggest that the LLMFPs face high barriers to market entry due to resource-intensive requirements for training models and potential legal challenges. Established players, like OpenAI and Anthropic, dominate the market, with proprietary models likely to lead to further consolidation. While some argue that open-source models foster competition, the trend indicates a move towards consolidation with limited new entrants. Regulatory bodies, like the Federal Cartel Offices, could play a crucial role in shaping this trajectory. Conversely, LLMLPs have lower barriers to entry, facilitated by APIs from LLMFPs. However, they face the risk of obsolescence as LLMFPs integrate similar functionality or discontinue APIs. Despite the regulatory landscape, LLMLPs face less legal challenges as they mostly work and train their refined models with available customer data.