Deployment gravity. Every market eventually splits in two: a premium tier that competes on being the best, and a workhorse tier that competes on being everywhere. Software did it — premium SaaS above, commodity Linux below. AI is doing it now, and the tier most people are watching is the wrong one.
The numbers behind that split landed in one week. Chinese open-weight models — versions whose trained parameters anyone can download and run — accounted for 41% of downloads on Hugging Face this spring, surpassing U.S. models. The top six models on OpenRouter are open-weight models from Chinese firms; Claude Opus 4.7 trails in seventh. Vercel, a developer platform that routes AI traffic for production apps, reported that open-weight models handled about a third of its AI requests in June.
The mechanism is repeatable. When production cost outweighs peak performance, ownership beats rent. Companies that scale on closed frontier APIs start watching inference bills the way they once watched AWS tabs — and reach for models they can host, fine-tune, and keep. The frontier stays a premium layer; deployment gravity pulls everything else down-market.
The capability race still runs on closed frontier models. The deployment race is running somewhere else entirely.
Reported by Sky for Type0, from The real AI race may no longer be at the frontier. Read the original: techcrunch.com