Five supply models, not one market.

The question that matters is no longer who has the most hours. It is which of the five models a buyer is actually contracting with, because they differ in what can be licensed, how fast data ships, and whether it can be used commercially at all.

For informational purposes only — see the disclaimer at the end of this note.

1 · What changed

Through 2025, this looked like any other data market: more hours cost more money, and the biggest dataset won. That ended when permissively licensed supply caught up with the size of a large commercial corpus. At the 100K-hour scale, hours alone stopped being a moat.

Build AI released Egocentric-10K in November 2025 and Egocentric-100K a month later. The larger release is 100,405 hours under Apache 2.0 — free for commercial use, and verifiable on the dataset card. Both are gated on Hugging Face: users share contact details to get access. Free first-person video is not new; Ego4D has long offered about 3,700 hours under a signed license. What is new is free supply at commercial scale.

Sources: Ego4D project documentation; Build AI and Lightwheel dataset cards on Hugging Face; Lightwheel’s release note of August 26, 2026. All accessed September 19, 2026.

Free is not the same product as paid. The 100K release ships at 256p with no IMU, no hand pose and no contact labels; the 10K release is 1080p but an order of magnitude smaller. At the 100K-hour scale, free supply is still raw factory footage. The annotated tier is catching up: in August 2026, Lightwheel began publishing EgoSuite-Open100K — 10,000 hours live out of a planned 100,000, spanning homes, hospitality, retail, sports, logistics, offices and industrial sites, with hand and body pose, licensed for commercial training but not for resale. The claim here is narrow, and hard to argue with: a paid hour now has to be different from a free hour — not just more of the same.

None of this means buyers are spending less. Volume is still valuable; generic volume is no longer scarce. Dyna-2 was pre-trained on more than one million hours of human video, and the company describes a path to ten million (Dyna Robotics, August 10, 2026). Figure’s in-house Index app was taking in 30 minutes of video every second when it launched on August 25, 2026, and the company says it will spend more than $1 billion on data and compute over the following twelve months. Micro1’s CEO estimates that robotics companies now spend more than $100 million a year buying real-world data (MIT Technology Review, April 1, 2026). One vendor was projecting a $100 million annual run rate by the end of 2026 (TechCrunch, September 11, 2026). And a 2026 Sinolink Securities report puts all usable, high-quality real-world interaction data at roughly 500,000 hours — far short of what general-purpose models are thought to need. Supply is elastic: all it takes is a head-mounted camera and an hourly wage. So strong demand shows up as volume, and as a premium on whatever is hard to copy. It does nothing for the price of footage anyone can record.

2 · Five supply models

Strip away the vendor names and there are five ways to get an hour of egocentric data. They produce legally and technically different assets. One question separates them: who signs, and what can they actually grant? Price, labor model and annotation depth all follow from the answer. A vendor can sit in more than one bucket: Build AI captures inside factories and releases openly; Lightwheel sells annotated hours and now gives some away. Teleoperation and purpose-built collection sites produce a different unit and are out of scope for this note; they appear below only for reference.

These are not quality tiers. A gig network is the right source for public-space navigation data and the wrong one for a commercial kitchen. An open dataset is the right place to start a visual backbone; it is not, on its own, something to build a deployment on. Ranking all five in a single list hides the one distinction that actually decides a purchase.

3 · Demand and capacity are in different places

This market is often described as consolidating around China. That is true of capacity and false of revenue, and conflating the two makes for bad forecasts.

The buyers that vendors actually name — frontier labs and robot makers — are concentrated in North America. Operated collection capacity is concentrated in Asia: purpose-built collection sites in China, and the lowest published contributor rates in India. Capital has followed both. In China, Lightwheel alone closed three rounds between March and June 2026 (Xinhua Finance, June 25, 2026); by the third quarter, the largest reported financings in progress were in the United States (XDOF, Mecka; TechCrunch, September 4 and 11, 2026).

United StatesDoorDash Tasks, up to · Washington Post, Apr 17, 2026

$25 / hour

Global contractorsMicro1 · MIT Technology Review, Apr 1, 2026

$15 / hour

Chinahousehold recording · Rest of World, Jun 3, 2026

¥20 ≈ $3 / hour

Indiareported rate · Bloomberg, Aug 12, 2026

≈ $2.62 / hour

Indiabase rate · TechCrunch, May 26, 2026

$1 / hour

Rates are not directly comparable: coverage, equipment and QA differ by market. Current job postings span a similar range, from $3 per accepted recorded hour in Southeast Asia to $30 an hour in North America (Chemin, July 2026; OpenTrain AI, September 2026). For reference, and not a wage: real-robot (teleoperated) data in China is priced at roughly ¥500–1,000 per hour (Nomura, July 2026, as reported by Tencent News, September 12, 2026). The order of magnitude is the point.

The implication is simple. A vendor whose costs and revenue sit in the same high-cost region is competing against vendors that have split the two — and against free. The models that survive that math either collect where costs are low and sell where demand is, or sell something that is not priced by the hour.

Splitting cost from revenue is not free, either. Footage of identifiable people leaving China falls under the PIPL and the 2024 cross-border rules: a standard-contract filing at lower volumes, and a security assessment once sensitive personal information of 10,000 or more people is exported in a year. India’s Digital Personal Data Protection Act applies to capture there. On the demand side, the FCC’s July 2026 decision to bar new foreign-made humanoid and quadruped robots is about devices, not datasets — but it shows how provenance questions are going to be asked. The arbitrage only works with a documented export path and per-engagement disclosure of where the data was collected.

4 · What became scarce

Undifferentiated video at the 100K-hour scale is no longer the bottleneck. That does not mean volume stopped mattering: NVIDIA’s EgoScale reports a log-linear scaling law across 20,854 hours of action-labeled egocentric video (arXiv:2602.16710), and vendors now report deliveries of human data in the millions of hours — company-reported figures, and not all of it egocentric video.

Figure 3 places the main datasets on two things that can be read off a dataset card: how many hours, and how much supervision those hours carry. It separates what the data contains from what has been demonstrated with it, and what has been released from what is merely held or planned. A million hours held, a million hours released and a million hours trained on are three different things.

released held or trained on, not released planned lift shown in a paper lift company-reported MOVAS, range across layers

Placements follow each publisher's paper, dataset card or official page, and are our classification. MOVAS-Ego 100K is drawn as a range because its layers are published but the hours carrying each layer are not; its contact layer is inferred from hand-surface distance, and no evidence of lift has been published. Maxinsights is placed according to its own description of tactile and force layers. Lift shown in a paper: Ego4D via R3M (CoRL 2022); EgoVerse (RSS 2026); EgoScale (arXiv:2602.16710). Dyna-2's scaling result is company-reported.

What separates a paid hour from a free one is a short list of four inputs. None of them comes from pointing more cameras at more people.

Rights that transfer. An individual can consent to the use of their own likeness. They cannot waive their employer’s confidentiality over the processes visible in frame, grant access to the site, or consent on behalf of the coworkers who appear alongside them. Most employment agreements restrict on-site recording before consent even comes up. A rights chain that stops at the contributor cannot clear what workplace footage requires, however carefully it was gathered. Nor can an employer’s signature stand in for the individuals’: it opens the site and covers confidentiality, and each person in frame still consents for themselves. A copyright license is not a rights chain, either: Apache 2.0 governs copying and is silent on privacy, likeness, biometrics and trade secrets.

Signals captured at the source. Labels derived from video — hand pose, camera trajectory, estimated contact — can be recomputed by anyone as open perception models improve. Streams that were never recorded cannot: unstabilized ego-motion, inertial data, depth, tactile and force. Vision alone does not recover weight, friction or contact force, which is why several vendors now pair head-mounted capture with sensorized gloves. What is sensed at capture time is a more durable advantage than what is annotated afterward. MOVAS-Ego 100K records stereo video, hardware-synchronized inertial data and unstabilized ego-motion; it does not include tactile or force streams.

Refinement into robot-native data. Raw video is already supervision — world-model and latent-action pre-training use it directly. It is not yet robot-native supervision. Hand-surface contact state, metric-scale geometry and retargeted trajectories validated against a specific end effector are the difference between a dataset and a policy input. Open releases now carry hand and body pose; none at scale yet carries contact state, metric-scale geometry or retargeted trajectories validated against a specific end effector.

Evidence of lift. This market has almost no independent evaluation. Buyers are asked to take it on faith that a dataset improves a policy, and the vendor selling the data is usually the one measuring it — ourselves included; see Method and disclosures. Neutral evaluation is the most obviously missing layer, and the likeliest to appear next.

5 · The comparison

Vendors and sources of egocentric hours, one row each. One rule: every row carries a source, and no row gets in without one. The Model column is our own classification, per Section 2. Scale is stated in each provider’s own accounting and labeled as such — held, delivered, released, uploaded or planned — because those are different quantities. MOVAS AI is listed as one vendor among several and held to the same standard.

“Not published” means the provider does not state it publicly — not that it does not exist. “Company-reported” means published by the provider and not independently verified. Hours are in each provider’s own accounting and are not a common unit: inventory held, hours delivered, hours released and hours planned are different quantities, and Lightwheel describes its 1.5 million hours as human data, not egocentric video specifically.

Adjacent: robot-native

XDOF. Teleoperation plus wearable-sensor human capture; in-house pipelines; about 20 customers, including frontier AI labs. Commercial hours are not published. Its open release, ABC-130K, is 134,806 episodes across 195 tasks, 3,553 hours, under Apache 2.0 (arXiv:2606.27375). Robot-native trajectories are a different unit from human video and sit outside the five models. Sources: SiliconANGLE, Jun 17, 2026; TechCrunch, Sep 4, 2026.

Left out of the table

6 · What to watch

Five things would change this analysis, roughly in order of likelihood.

Sources

1. Build AI, Egocentric-100K and Egocentric-10K dataset cards and organization page, Hugging Face — accessed September 19, 2026.
2. Lightwheel, EgoSuite-Open100K release note (Hugging Face, August 26, 2026) and EgoStandard dataset card — accessed September 19, 2026.
3. Figure, “Introducing Index,” August 25, 2026; Figure and Brookfield partnership announcement, September 17, 2025.
4. Washington Post, April 17, 2026 (DoorDash Tasks); MIT Technology Review, April 1, 2026 (Micro1); Rest of World, June 3, 2026 (China); Bloomberg, August 12, 2026 (India).
5. Ego4D and Ego-Exo4D project documentation (license scope, hours, annotations) — accessed September 19, 2026.
6. Dyna Robotics, Dyna-2 announcement, PR Newswire, August 10, 2026.
7. TechCrunch, “Mecka AI nears $500M valuation in Sequoia-led deal,” September 11, 2026; and its report on XDOF, September 4, 2026.
8. TechCrunch, “This startup is betting India’s gig economy can train the world’s robots” (Human Archive), May 26, 2026, citing the Economic Times and Moneycontrol.
9. SiliconANGLE, “Robotic teleoperation data startup XDOF launches with $70M in funding,” June 17, 2026; Allshire et al., arXiv 2606.27375, June 25, 2026 (ABC-130K).
10. Xinhua Finance, interview with Lightwheel, June 25, 2026; Zhongguancun Science City notice, June 1, 2026.
11. Tencent News, September 12, 2026, citing Nomura (China robotics report, July 2026) and Sinolink Securities (China embodied-intelligence industry report, 2026).
12. OpenTrain AI contributor postings, September 4 and 10, 2026; Chemin contributor posting, July 2026.
13. FCC, Covered List update adding foreign-produced advanced robotic devices, July 28, 2026.
14. Cyberspace Administration of China, Provisions on Promoting and Regulating Cross-Border Data Flows, March 22, 2024, Articles 7–8.