technologyScore 55/100Research
Karpathy notes LLMs have properties of semiconductor fabs: high CAPEX, centralized R&D secrets, and defensible technology trees
Andrej Karpathy· Independent· AI· 2026-06-03· about OpenAI, Google, Anthropic
“LLMs don't only have properties of utilities. I think it's also fair to say that they have some properties of fabs. And the reason for this is that the CAPEX required for building LLMs is actually quite large... You're investing a huge amount of money, and I think the tech tree for the technology is growing quite rapidly. So we're in a world where we have sort of deep tech trees, research and development secrets that are centralizing inside the LLM labs.”
Why it matters
LLM development is characterized by high barriers to entry, rapid technology advancement, and concentration of R&D talent and secrets in a few labs. This mirrors semiconductor fab economics.
Investment implication
Smaller players and open-source alternatives (like Llama ecosystem) may struggle to keep pace without access to similar CAPEX and talent. Companies enabling LLM development tooling or optimization (inference, fine-tuning) become critical picks-and-shovels plays.