Open Models Approaching 3T Parameters; Inference Compute Might Be the Only Remaining Moat

burkov · x · 2026-07-21

The author points out that current open-weight models are now close to 3 trillion (3T) parameters, and Fable is rumored to have 6-8T parameters. Following this trajectory, we might see colossal models with 20T parameters or more by the end of the year.

Consequently, they suggest that the ability to efficiently run inference for such massive models might become the only remaining competitive moat for AI companies.

Related event: Open-Source Models Approach 3T Parameters, Inference Compute Becomes the New Moat(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →