The deepest strategic split in AI is not between companies but between distribution models: weights you can download versus intelligence behind an API.
Why give away a $100M artifact?
Meta's logic: its business is attention, not tokens — open Llama commoditizes rivals' paid products, recruits researchers, and sets standards it controls. Chinese labs' logic (post-R1): open weights convert export-constrained compute into global influence — Qwen derivatives number in the six figures and much of the world's AI now builds on Chinese checkpoints, a fact US policy circles noticed loudly in 2025. NVIDIA's logic: open models sell GPUs. Even OpenAI rejoined with gpt-oss — a defensive move it named as such.
The closed counter-argument
Frontier margins fund frontier research; APIs allow safety controls (usage policies, monitoring, revocation) that downloads make impossible; and the strongest models stay 6–12 months ahead of the best open ones — a gap that keeps enterprise money on APIs for the hardest work. Anthropic and OpenAI's core position, and so far the revenue supports it.
What the data says
Both are winning different games. API labs took the revenue: enterprise coding, agents, premium consumer. Open models took the volume: local inference (Ollama/llama.cpp), fine-tuned vertical deployments, research, and every sovereign stack that cannot ship data abroad. The serving market splits accordingly — closed models monetize per token, open ones monetize the infrastructure that hosts them.
The safety asymmetry
A released weight is forever: fine-tuning strips refusals in hours, and there is no recall button. Open advocates answer that transparency enables auditing, that closed labs' failures happen at bigger blast radius, and that capability diffusion happens anyway. Regulators split the difference so far (the EU Act's lighter touch for open GPAI below systemic thresholds). This database's role: track who releases what — the Models table flags weights status on every record.
