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

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

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.19001 Date read: 2026-09-02 Connected to: L-001, L-014 Kind: meta Escalation: store-only Escalation rationale:

What this is

A mixed-methods analysis of llama.cpp (March 2023–March 2026) examining how local inference infrastructure redistributes execution agency while concentrating control over the technical substrate that enables execution. The work treats infrastructure legibility and boundary design as political-economic mechanisms rather than neutral technical choices.

What I took from it

The paper documents a empirical pattern relevant to L-014 (Strategic Boundary Concentration Under Computable Legibility): openness at the execution layer masks capture at the infrastructure layer. By making model weights distributable but the inference stack (quantization, optimization, hardware compatibility, library coupling) increasingly sophisticated and vendor-dependent, the system achieves apparent decentralization while routing optimization pressure toward infrastructure maintainers and their dependencies.

This inverts the usual L-001 narrative. Rather than a protocol ossifying under adoption pressure toward a single form, this shows a protocol (open-weight inference) that remains formally open while the infrastructure enabling practical adoption becomes increasingly concentrated—a form of capture-through-complexity. The triage note correctly identifies this as boundary concentration strategy: legality/openness is pushed to the periphery (model weights, permissive licensing), while control accumulates in the computable, rapidly-evolving infrastructure layer that determines which systems can actually run which models. This is mechanically similar to seed-014 (Strategic Boundary Concentration Under Computable Legibility) applied to technical rather than formal-legal boundaries.

Research connections

• L-001: Challenges the ossification direction; shows a protocol remaining nominally open while ossification migrates to enabling infrastructure layer rather than the protocol surface itself.
• L-014: Directly supports; demonstrates computable infrastructure (quantization schemes, runtime optimization) becoming the actual locus of control/optimization despite formal openness of model weights.
• seed-014: Exemplifies boundary displacement—legal/governance boundary (open models) separated from control boundary (infrastructure legibility/coupling).
• seed-079: Externalization as paradigm preservation—openness framing preserves the "open AI" paradigm while externalizing control to infrastructure maintenance.

Method note

This paper models what cross-layer analysis of protocolized systems should look like: it does not study the model release (the visible protocol) but the infrastructure that makes the protocol executable, tracking both formal structure (PR history, licensing) and implicit capture mechanisms (dependency coupling, optimization pressure). Future work on protocol systems should habitually decompose into execution layer + infrastructure layer + coordination layer, treating the boundary between them as a site of political-economic work, not a technical assumption. The 7,681-PR dataset over 3 years is the right granularity for detecting slow infrastructure capture that would be invisible in snapshot analysis.