Source: Alex Heath
The economics of training large language models—requiring massive compute, data, and capital—concentrate power among a handful of companies (Anthropic, OpenAI, Google, Meta) who can afford the infrastructure, leaving smaller labs and academic teams dependent on renting API access under terms those companies control. This shift from open-source to closed access matters because the companies controlling foundational models also control what research questions get asked, what safety constraints get embedded, and who can compete in downstream applications. The open research movement's risk isn't losing altruism—it's losing the ability for anyone outside these walls to audit, modify, or contribute to the systems reshaping knowledge work and AI policy.