Atlas sold out its production year — and the humanoid market split in two

Boston Dynamics has committed all of 2026's output before most of it exists. The constraint is no longer capability; it is manufacturing slots, and that has cleaved the industry into a volume tier and a capability tier.

Electric Atlas units are shipping to Hyundai's metaplant and Google DeepMind, with the entire 2026 production run already spoken for. Selling out a year in advance settles the capability debate by making it irrelevant. The question now is who can build and deliver.

Two customers, two purposes

The destinations are chosen, not incidental. Hyundai is Atlas as industrial equipment inside its corporate parent's factory — proving the business case. DeepMind is Atlas as a research platform for embodied learning — feeding the models. One earns revenue, the other generates the manipulation data the next generation trains on.

The split

Against Figure past a thousand units at one per hour, AgiBot at fifteen thousand cumulative, and Unitree targeting up to twenty thousand at a tenth of Western prices, Atlas is deliberately the low-volume premium end. The market has stratified into a volume tier riding a cost curve and a capability tier selling what cheap platforms cannot reliably do.

What decides it

Probably neither price nor polish, but data — and specifically touch. A single policy now spans whole-body teleoperation, learned locomotion, dexterous hand retargeting and distributed tactile sensing. Most manipulation failures are contact failures, and contact is invisible to cameras. Whoever accumulates the most hours of real touching wins the part of the problem that vision cannot reach.

Drones and EVs both ran this bracket. The volume tier won both times. The variable this round is whether embodied capability scales with fleet-hours the way language scaled with tokens.

IEEE Spectrum — Video Friday: humanoid robot production → · Humanoid Press — Humanoid Daily — robotics and AI news →