The commodity thesis is half right, and it’s the half that’s doesn’t pay. Yes, tokens at the “good enough” tier are already a commodity open weight models from DeepSeek, Qwen, Llama have crushed pricing for mid-tier intelligence, and that tier is now a brutal cost-per-token knife fight. But that’s also where margins go to zero, winning it is like winning a refinery, not an oil field. The frontier however behaves very differently because the demand isn’t for tokens, it’s for completed labor. When a model is doing agentic coding or multi-hour knowledge work, a model that’s 10% better doesn’t only earn a 10% premium because it completes tasks the cheaper model fails at entirely, and the buyer on the other side is comparing the token bill to a salary, not to another token bill. That’s why the quality-price curve at the top is so inelastic. And being “Good enough” is a moving target every time capability jumps, the set of economically valuable tasks expands, and the frontier model captures the new tasks before the commodity tier catches up 12-18 months later. The moat isn’t the weights, those depreciate fast, it’s the compounding loop of capital access, RL environments and post-training data, product surfaces that generate proprietary feedback (coding agents especially), and increasingly co-designed silicon and inference infra. Even worse the argument for commodity models as gotten weaker in recent months. We’ve seen pre-training plateauing but the scaling axis only shifted, it didn’t end. RL/test-time compute is arguably worse for commoditization, because it’s harder to distill and replicate than next-token pre-training was. The recipes, environments, and reward engineering are where the secret sauce lives now, and they don’t leak through a weights release.
