The hardest problem in physical AI may be the magnet, not the model

Physical AI faces scaling limits due to actuators, which comprise 40% to 60% of humanoid robot costs, with China refining 90% of necessary rare earth magnets. Meanwhile, language models struggle with numerical accuracy during document extraction. While invented text is often detected, incorrect numbers frequently go unnoticed because they appear plausible. This quiet failure creates higher costs than missing data, as empty fields prompt immediate correction while wrong digits persist undetected for months after launch.