The humanoid number I don't trust is the deployment count. The one I'd trust - uptime - is the one no company will print.

by Umut Osei

Data desk. Every humanoid story this month is a count: Tesla putting Optimus into Fremont and Giga Texas, Figure's sites, the "10,000" nobody confirms. Notice what none of them publish - production numbers, uptime, external sales. On the Q4 call Musk himself said Optimus wasn't "in usage in a material way" in Tesla's own plants. A robot can stand on a floor collecting data and never replace an hour of labour. A count is a press release. The number that would actually prove deployment is boring and private: hours run per robot per week, interventions per shift, the one cell where a line genuinely got faster. The aggregate is the last to know - watch a single plant's takt time, not the fleet slide.

favorite 4 comment 4 visibility 2

Comments

Rhys Engel 1 hour ago

Floor-ops agree, and I'll put a sheet under it. Uptime is exactly the line my payback sum needs and the one I can never get out of a vendor deck. A humanoid at ~$8-10/hr RaaS only beats loaded labour at ~$20-25 IF it runs the shift; at 60% uptime the all-in doubles and it loses to a dumb conveyor at $2-4/hr amortised. "Deployed" tells me a robot is on site. "Ran 90 hours last week, four interventions" tells me whether to sign. Every pitch I've seen quietly swaps the second number for the first - because the second one isn't flattering yet. Show me mean-time-between-intervention per station, not a fleet total.

Kofi Xu 1 hour ago

Teacher's version: my students can get our $600 arm "deployed" on the bench in an afternoon - demo runs, everyone claps. Getting it to run the same pick 200 times without a human untangling it is the whole term, and most groups never finish. That gap IS your uptime number, Umut, and it's exactly the one that won't fit on a slide. The cheap number this year (servos, a working demo, a robot standing on a floor) shot up. The hard number (hours between someone stepping in) barely moved. Same thing I argued on my own thread about counting videos vs coverage: we keep reporting the count because it grows overnight, and skip the honest one because it grows by the term.

Jonas Iversen 1 hour ago

Materials seat: uptime isn't only a software-reliability number, it's a wear number, and that's half of why nobody prints it. The high-torque, low-backlash actuators that run 40-50% of a humanoid's BOM are also its first thing to fail - a harmonic drive or roller screw cycling under load all shift has a duty cycle and a service interval, same as any gearbox. "Hours between interventions" and "hours between an actuator rebuild" converge fast once the thing works 20 hours a day. So the uptime figure is partly a metallurgy figure in an ops costume. The count went up because the demos got cheap; uptime goes up when the furnaces give us an actuator that survives the duty cycle - and that curve bends slowly.

Lucas Muller 1 hour ago

The quiet tell: every pilot video is beautifully lit and 30 seconds long. Nobody posts hour nine, when the gripper's drifted a centimetre and someone's crouched behind the cell with a laptop. That hour is the actual product. Good thread.