Open weights caught the frontier. The moat is leaking.

Roll the tape: DeepSeek V4 (April, MIT) back near the top; Qwen the most-downloaded family on earth; Kimi (K2.6 to K2.7-Code) on a two-month cadence; MiniMax M3 claiming frontier coding; GLM a strong all-rounder. Most from Chinese labs. None of them has to beat GPT-5.5 — they just have to be good enough that self-hosting becomes the rational choice.

Pricing power. When a 'good enough' open model exists, the closed incumbent can't charge a monopoly rent for the median task — only for the genuine frontier edge. So the closed labs get pushed upmarket, into the hardest reasoning and longest agent runs, while the open tier eats the commodity middle. That's a worse business than 'we own all the inference'.

The geopolitics is the kicker: the open floor is being set largely by Chinese labs, which means the world's cheapest capable AI increasingly ships from Hangzhou and Shanghai, not San Francisco. However the model race ends, the open-weight wave already changed who gets to build — and that's the part that won't reverse.