Open at One Layer: Licensing, Data Provenance, and Hardware Costs in Global AI Deployment
Year: 2026
Type: preprint
Abstract: Openness in artificial intelligence belongs to layers. By treating publicly released numerical parameters as proof of openness, advocates promote foundation models to low-income countries while leaving training data and local hardware costs unexamined. Working deployments depend on all three layers. Both the Open Source AI Definition and European Union law mandate public parameters alongside summarized training records. In licensing, Meta's Llama agreements bar unlicensed medical practice under California jurisdiction, whereas Google removed usage restrictions from Gemma 4 in 2026 under the Apache 2.0 license. Training records invert the licensing divide. Foreign laboratories withhold training data, while African teams releasing regional corpora adopt protective terms that international standards classify as closed. Across 2026, data-center demand inflated smartphone memory prices, raising retail costs for budget handsets. Local deployers remain bound by the most closed layer. Keywords: Open-Weight AI, Foundation Models, Technological Leapfrogging, Open Source AI Definition, EU AI Act, Digital Public Goods, Llama Community License, Apache 2.0, Pretraining Data, Training Data Provenance, Low-Resource Languages, Small Language Models, InkubaLM, AfroLlama, Handset Affordability, Memory Chip Inflation, Silicon Supply Chain, Edge Inference, Digital Divide, Sub-Saharan Africa
Authors: Jefferson, Josie, Velasco, Felix
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236080319
