Nomura: Embodied AI video data prices fall, while contact data retains a near-term premium

Why are video prices falling as data purchases grow?

Nomura’s industry survey indicates that a leading US laboratory spent approximately 1 billion US dollars on data in 2026 and plans to spend approximately 5 times this year’s amount in 2027. But a larger procurement budget does not mean prices can rise for every type of data.
First-person video has relatively low collection barriers, and new supply has pushed prices down: the survey found that prices were approximately 40–50 US dollars per hour at the beginning of 2026, but had fallen to near or below 10–15 US dollars by the time of the report. Nomura expects such raw video to move gradually toward cost-plus pricing.
The shares of hours and spending also differ. Nomura estimates that first-person video accounts for about three-quarters of the laboratory’s target hours; teleoperation data accounts for only a mid-single-digit percentage of hours but more than a quarter of spending.

Video primarily helps models learn perception and understand the world; teleoperation records robots’ actual movements, including force and contact signals, and can be used to train finer manipulation skills.
Nomura estimates that the median hourly price of raw teleoperation data is more than 7 times that of first-person video. But only 20%–30% of purchased video is usable after screening, versus approximately three-quarters of teleoperation recordings; measured by usable hours, the teleoperation premium narrows to approximately 2–4 times.
Supply expansion is constrained by actual collection capacity. Nomura’s survey indicates that an in-house teleoperation site with 45–55 sets of equipment, operating three shifts year-round, produces approximately 250 thousand hours of usable data annually. At that rate, reaching the laboratory’s target would take several years, and most demand would still need to be met by external suppliers or larger sites.
Equipment reliability is also a constraint: haptic gloves still have high failure rates and have not yet been deployed at scale. Nomura estimates that collection devices costing approximately 4000 yuan each often break within two or three months.

Nomura believes data integration, validation and contact-data collection closer to buyers offer greater pricing room, while margins on field collection alone are thin. Using low-cost raw video from Southeast Asia as an example, the bank estimates that a tier-one supplier retains approximately 5 US dollars per hour, while collectors receive 1–2 US dollars.
Post-processing can increase data selling prices. A laboratory case in Nomura’s survey showed that post-processing raised the hourly price of first-person data from approximately 60 yuan to 150–200 yuan. The bank also noted that processing capabilities, automated annotation and customer access are more defensible than field collection.
The room for Chinese suppliers to enter the core supply chains of leading US laboratories is narrowing. Nomura’s survey says the laboratory excludes suppliers with Chinese ties in its phased tendering process; China-related businesses therefore rely mainly on domestic buyers and indirect overseas channels.

Video’s ability to replace teleoperation is improving. The EgoScale preprint uses 20854 hours of human video for pretraining, with mid-training requiring only approximately 50 hours of aligned human data and 4 hours of robot data. Based on the relevant paper, Nomura believes video could replace most teleoperation data in pretraining in the first half of 2027.
However, Nomura expects comparable substitution in post-training involving rich contact may not occur until after 2028. Its recent industry survey estimated a task success rate of approximately 60%, while home use requires 98%–99%; contact data will retain value until that gap narrows.
Nomura would reassess data premiums and the distribution of value across the supply chain if video’s usable yield after screening rises substantially above the current 20%–30%, teleoperation prices fall, or haptic glove reliability improves.