Global Robotics — Physical AI data: volume up, prices down

Following the report Global Robotics - The ChatGPT moment comes into view published on 23 October, this report looks at data collection and argues that frontier physical-AI programs buy volume cheaply and pay up for precision in the near term. Based on analysis of industry data and field survey, egocentric video accounts for roughly three-quarters of the lead US lab's target hours, while teleoperation is a mid-single-digit share of hours but over one-quarter of the spend. Only 20-30% of purchased video survives filtering, versus three-quarters of teleoperation recordings. Ego data fell from USD40-50 per hour early in 2026 to close to or under USD10-15 now, as low entry barriers drew in supply. The lead lab's data spending this year is USD1bn, with about 5x that planned for next year. Contact data stay scarce, and incremental spend should favour precision; teleoperation and force data keep a premium because supply cannot scale quickly. Raw overseas ego data trade at USD10-35 per hour be