Open at the Edge, Captured at the Center: llama.cpp and the Political Economy of Local AI Inference

Research into the political economy of local AI inference, specifically through the case study of llama.cpp, reveals a trend where broad participation at the execution level is accompanied by a concentration of control within the underlying infrastructure. While open-weight models allow for local operation, critical infrastructure components such as hardware backends, model integration, and distribution platforms are becoming points of 'capture' by specific vendors and maintainers.

Why it matters

This analysis highlights an emerging dynamic in the technology sector where open-source initiatives, while promoting widespread access and innovation, can paradoxically lead to new forms of centralization and control. Organizations must understand these shifts to strategically navigate the development and deployment of AI, particularly concerning vendor dependencies and the long-term viability of decentralized solutions.

Key insights

• Local AI inference, exemplified by projects like llama.cpp, significantly broadens user participation in running open-weight AI models on personal devices.
• Despite decentralized execution, control over the local AI ecosystem is shifting to hardware vendors and model distributors.
• The absorption of projects like llama.cpp by larger entities, such as Hugging Face in February 2026, indicates a centralization trend in the infrastructure layer.
• Analysis of 7,681 merged pull requests over three years, repository discussions, corporate statements, and contributor blogs formed the basis of this mixed-methods study.
• This phenomenon is termed 'capture at the center,' where infrastructure providers gain significant influence over the distributed local inference landscape.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.19001