Dreamscale serves NVIDIA's DreamZero robot model 8% faster with half the silicon

chris_j_paxton · x · 2026-09-23

Dreamscale Labs (YC F26) claims to be the first inference provider serving NVIDIA's DreamZero robot model over the cloud, beating NVIDIA's own reported inference speed by 8% with half the silicon. DreamZero pioneered the "World Action Models" wave in generalist robotics, promising better task generalization, learning physics from human video, and transfer to new embodiments. The catch: NVIDIA's setup needs 2x GB200 superchips ($60K each) or is painfully slow (5.7s wait for 1.6s of motion), making real-robot deployment infeasible — Dreamscale's real-time cloud inference claims to fix that.

Original post →

More from Embodied

Embodied channel →