Open-weight models may take the largest share of global AI usage
joshua_saxe · x · 2026-07-24
The author argues that US policymakers should worry less about blocking Chinese training on US model outputs and more about the broader shift toward open-weight models.
Key claims:
- Countries pursuing sovereign AI will prefer customizable local models, and many of those open-weight options are coming from China.
- Open-weight usage appears to be on a strong growth curve, with OpenRouter traffic showing open models overtaking closed ones.
- Inference providers focused on open weights will keep pushing costs down aggressively, which could matter as tokens become commoditized.
- A horizontally decoupled ecosystem — training, inference, harnesses, and apps split apart — may outcompete vertically integrated stacks.
The attached chart claims open weights moved from 19% to 60% of OpenRouter traffic in under seven months.
Related event: Open-Weight Models Set to Dominate Global AI Usage(3 posts)→
More from AGI Musings
- Sam Altman says OpenAI wants an AI research intern on hundreds of thousands of GPUs by 2026 — morgymcg · 2026-07-24
- Teachers should adapt to LLMs instead of treating them like cheating — NickPassig · 2026-07-24
- Code review may disappear by 2027, but production code still needs human eyes — tlakomy · 2026-07-24
- A satirical AI-disempowerment joke turns safety talk into a punchline — jd_pressman · 2026-07-24
- AI incidents may change regulation debates, but not prove a bigger disaster is near — binarybits · 2026-07-24
- Miles Brundage warns AI still risks sliding back to business as usual — Miles_Brundage · 2026-07-24