# Robot Data Startups Now Claim $150 Million a Year. The Top Humanoid Maker Booked $250 Million.

> Markdown twin of https://roboterminal.app/posts/robot-data-sellers-vs-unitree · https://roboterminal.app/posts/robot-data-sellers-vs-unitree.md

**Published:** 2026-10-12

Mecka raised $60 million in a Series B led by Sequoia Capital on October 7 and says it passed $100 million in run-rate revenue in June 2026, with a projection of $300 million by year end.

Mecka says it passed $100 million in run-rate revenue in June and XDOF was reported near $50 million, together about 60 percent of what Unitree, the world's top humanoid seller, booked in all of 2025. The first physical-AI layer with real revenue sells recordings of people doing chores, and it is paid for by labs that may decide to collect their own.

**Companies:** [Mecka](https://roboterminal.app/companies/mecka), [XDOF](https://roboterminal.app/companies/xdof), [Scale AI](https://roboterminal.app/companies/scale-ai), [Micro1](https://roboterminal.app/companies/micro1), [Unitree](https://roboterminal.app/companies/unitree)

**Tags:** physical-ai, robot-data, humanoids, funding

[Mecka](https://roboterminal.app/companies/mecka) raised $60 million in a Series B led by Sequoia Capital on October 7 and says it passed $100 million in run-rate revenue in June 2026, with a projection of $300 million by year end (https://www.mecka.ai/news/series-b). [XDOF](https://roboterminal.app/companies/xdof), three months out of stealth, was reported in September to be approaching $50 million in annualized revenue and in late-stage talks to raise at about a $1.2 billion valuation (https://techcrunch.com/2026/09/04/xdof-just-three-months-out-of-stealth-is-in-talks-for-a-series-b-at-a-1-2b-valuation/). Run-rate revenue means current monthly sales times twelve. It is a company's own extrapolation, and every figure in this paragraph is either company-claimed or source-reported, not audited. Put them together and two sellers of robot training data claim about $150 million a year (arithmetic, two different dates, two different sources). [Unitree](https://roboterminal.app/companies/unitree) was the world's top humanoid seller in 2025, with roughly a third of global humanoid sales. It booked RMB 1.71 billion ($250 million) for all of 2025, and that figure includes its quadrupeds (https://restofworld.org/2026/unitree-china-humanoid-robot-shanghai-ipo/). Two companies that sell recordings of people folding laundry and fixing cars claim about 60 percent of the top robot maker's year (arithmetic). If Mecka hits its year-end projection, one data seller alone passes Unitree's 2025 revenue, at about 1.2 times (arithmetic).

The money in humanoids right now is in teaching them, and the raw input, in Interact Analysis analyst Marco Wang's description, is a recording device plus wages of $5 to $20 an hour paid to people doing chores on camera (https://www.cnn.com/2026/04/04/tech/humanoid-robot-training-jobs-intl-hnk-dst). Two terms matter for the rest of this post. Teleoperation is when a person steers a robot from a distance and the robot records the result. Egocentric data is video and sensor readings captured from the point of view of the person doing a task, through glasses, a GoPro, or a phone, with no robot involved. Our view: this is the first layer of physical AI with real revenue attached, and it is also a cost line funded by labs, sold to a short list of unnamed buyers, and exposed both to cheaper substitutes and to customers who decide to collect their own. **The robots are students, and someone else is paying their tuition.**

## Who sells what kind of robot data

The table below sorts the vendors by what they actually record rather than by who raised the most. The products come in a few forms: a human wearing sensors with no robot present, a human driving a robot by teleoperation in a warehouse built for the purpose, and a human driving a robot on a live job site. We include the Føtex supermarket trial in the last category because, in practice, it is a data operation: "human operators guide every movement" (https://www.euronews.com/video/2026/10/11/denmark-tests-humanoid-robots-in-supermarkets-to-help-to-ease-staff-workload-at-night). [Scale AI](https://roboterminal.app/companies/scale-ai) and [Micro1](https://roboterminal.app/companies/micro1) are included because both started in language-model data and moved into robotics (https://techcrunch.com/2026/10/07/robot-data-startup-mecka-ai-nabs-60m-from-sequoia/). Read the fourth column, which is mostly empty.

| Company | Data type | Scale or revenue stated | Buyers named | Source |
|---|---|---|---|---|
| Mecka | Wearable human demonstration (body sensors plus smartphones) in homes and commercial sites; builds its own capture hardware; also post-trains models as a self-described "modern robotics integrator" | Over $100M run-rate in June 2026; projects $300M by year end; $60M Series B from Sequoia, NVIDIA, M12, Qualcomm Ventures, Samsung | None ("several of the top frontier robotics labs," "multiple Mag 7 companies") | Mecka; TechCrunch Oct 7 |
| XDOF | Teleoperation on deployed and general robots plus egocentric wearable capture; three-tier "data pyramid"; ABC-130K open set of over 130,000 episodes across 195 tasks | About $50M annualized (reported); $70M raised; about 60 staff; 20 customers; talks at about $1.2B | None (frontier AI labs, unnamed) | TechCrunch Sep 4; TechCrunch Jun 17 (https://techcrunch.com/2026/06/17/collecting-robot-training-data-is-dirty-unglamorous-work-some-ai-labs-are-already-paying-xdof-to-do-it/); XDOF blog (https://www.xdof.ai/blog/abc-130k) |
| Scale AI | Dedicated data-collection robots plus human demonstrations; San Francisco prototyping lab plus global contributors | Over 100,000 production hours | Physical Intelligence, Generalist AI, Cobot | Scale AI (https://scale.com/blog/physical-ai) |
| Micro1 | Home egocentric video from about 4,000 "robotics generalists" in 71 countries | Over 160,000 hours a month; at least 10 hours a week per worker | Not named | CNN |
| Hive Robots / Salling Group (Føtex) | In-situ teleoperation of Unitree-built humanoids in a live supermarket | Training since June 2026; target is independent night restocking by early 2027 | Salling Group is the operator, not a data buyer | Euronews |

One row names customers and one names a date for a robot working alone, and they are not the same row.

## What is actually in the box: hours of a human, and one vendor says half are usable

XDOF's own ranking of its product is the clearest description anyone has published. The most valuable data, in its three-tier pyramid, is teleoperation collected on the exact robot that will be deployed. Below that sits teleoperation on robots gathering more general data. At the bottom is egocentric data from people doing everyday tasks, for which XDOF plans to build its own wearable sensors. Mecka's pitch for why any of this is worth paying for is one sentence: "Motion, contact, force and geometry aren't on the internet. You can't scrape them, license them or buy them." So Mecka pays people to make coffee or fix cars while wearing body sensors and using smartphones.

The atoms of the product are cheap and the yield is poor. Wang puts the cost of human data at a recording device, a GoPro, Meta glasses, or a phone, plus hourly wages of $5 to $20 depending on region. His summary of what buyers want: "I don't want the robot doing the task. I want the people doing the task." Objectways founder Ravi Rajalingam says only about half of the footage submitted to him is usable, that some customers pay more for data from US households where the hourly wage can be as much as triple that of a worker in Vietnam or India, and that 90 percent of his customers are US-based. XDOF, aware that "data provision alone can be a dead-end business," also sells cleaning, tooling and annotation.

Here is the tension inside the table. XDOF's value ordering runs from on-robot teleoperation at the top down to egocentric video at the bottom. The volume ordering, at least in these sources, runs the other way: the tier XDOF ranks lowest, people filming themselves at home, is the one producing the most hours, with Micro1's 160,000 hours a month as the clearest example.

## The volume math: 32 times the open corpus every month, 520 years from "enough"

Scale AI says the relevant open datasets, DROID and Open X-Embodiment, offer about 5,000 hours of interaction data combined. Micro1's roughly 4,000 contributors send more than 160,000 hours of video every month. One month of Micro1 is therefore about 32 times the entire open corpus (arithmetic), and a year at that pace is about 1.92 million hours (arithmetic). That is the gap in one direction: a single vendor at the bottom of XDOF's pyramid out-produces the open datasets Scale AI calls relevant, every month.

The gap in the other direction is larger. Micro1's Arian Sadeghi says the field needs "probably billions of hours." At 160,000 hours a month, the first billion takes about 6,250 months, which is about 520 years (arithmetic). Apply Objectways' observation that only about half of footage is usable and the figure gets worse, though we will not compute a new one because the usable fraction almost certainly varies by vendor and task.

Two facts pull the other way. First, Nvidia reported in February that adding more than 20,000 hours of first-person video to robot training improved success on tasks like rolling T-shirts, sorting cards, unscrewing bottle caps and using a syringe by more than 50 percent. Useful gains arrive at twenty thousand hours, long before billions. Second, market research firms estimate data collection and labeling will grow about 30 percent a year to at least $10 billion by 2030, a forecast CNN attributes to unnamed firms and one we note without leaning on it. Nvidia reported gains at twenty thousand hours; the billion-hour figure is Sadeghi's estimate, and nothing in these sources tests it.

## Why the money is in teaching before the robots earn anything

Start with the robot maker, because its revenue is tuition money too. Unitree sold 5,500 humanoids in 2025, and more than 70 percent of them went to research and education. Humanoids were 51.5 percent of core revenue, and the average humanoid price fell from about RMB 593,400 ($85,000) in 2023 to RMB 167,600 ($25,000) in 2025 while gross margin stayed near 60 percent. Wang's read on China as a whole matches: most humanoids mass-produced there so far were bought for training and research, and the country has announced plans for at least 60 robot training centers. Global shipments approached 25,000 units in the first half of 2026, up 432.1 percent, with AgiBot above 8,600 (https://www.humanoidhub.ai/news/agibot-leads-global-humanoid-robot-shipments-in-h1-2026-idc-data-shows). None of these sources says how many of those units are doing paid work.

The performance numbers explain why. IFR research chair Alexander Verl says T-shirt folding success is "usually around 70 or 80%," which is not commercially viable for manufacturing partners. A robot that fails two or three folds in ten (arithmetic from Verl's 70 to 80 percent) has to be taught more, and teaching requires infrastructure most labs do not have. XDOF's Philipp Wu on why labs outsource: "You need a warehouse of hundreds of thousands of square feet with hundreds of robots. You need to maintain these robots, calibrate their physical parameters, and properly train operators."

The Føtex store near Copenhagen is the classroom version of the same thing. Hive Robots' humanoids, built by Unitree, have been training since June with a human operator guiding every movement. Føtex workers spend 40 to 50 percent of their time replenishing shelves, so that is the labor target. The robots work more slowly than people and need close supervision, and Danish law would have to change before robots operate autonomously among shoppers. The stated goal is independent night restocking by early 2027, and until that date arrives, Salling Group is buying lessons rather than labor.

## Objection docket

The run-rates are soft, and we agree. Mecka's $100 million is a June figure, XDOF's $50 million is a September report from people familiar with a fundraise, and neither is audited. Mecka also says the $100 million came "within only a few months of operations," which is a pace. Nobody has seen a full year of it. The $150 million is what two companies tell investors, and with every buyer unnamed it measures their pitch rather than a market.

Buyers will insource, and this is already happening at the top. Tesla trains Optimus in its own facilities in Fremont and plans to expand in Austin. Figure built its own data pipeline, the Index video app, and said outside vendors could not deliver throughput, diversity and quality at the same time, which we covered at https://roboterminal.app/posts/figure-index-nscale-training-loop. Mecka's answer is to call itself an integrator and post-train on the data it captures, and XDOF's is to sell cleaning and tooling. Both are hedges against the same risk.

Simulation or video conversion could make the whole layer cheaper. Labellerr AI co-founder Puneet Jindal says the human-data boom may not last if simulation improves or AI can convert YouTube video to first-person. His sharper line: "Even robotics labs are feeling like they don't know what data will be needed 12 months from now." We concede this is unresolved. The Nvidia result shows first-person video helps, and nothing in CNN's summary of it says whether the video must be freshly recorded.

Open data might close the gap. Not at current volumes. ABC-130K, over 130,000 episodes across 195 tasks on YAM arms under an Apache 2.0 license, is the largest open teleoperation set to date by XDOF's account, and it was released by a vendor together with UC Berkeley researchers, so it doubles as a sample of the paid product. We track open releases at https://roboterminal.app/posts/open-robot-models-download-test.

## Three numbers to check

1. **Mecka's run rate at the end of 2026.** The company projects $300 million. Anything near that would put one data seller past Unitree's whole 2025 revenue; a quiet miss would say the June figure was a peak.
2. **Whether XDOF's round closes, and at what terms.** The September report described talks at about $1.2 billion with terms not final. A closed round with a disclosed size turns a rumor into a price for this layer.
3. **Whether a Hive robot restocks a Føtex store without an operator by early 2027.** That is the date Salling Group's trial set for independent night work. It is the first point where a buyer in this table stops paying for lessons and starts getting labor.

Tuition is paid in advance and returned in wages later, or not at all. Right now the labs are paying it, the data vendors are collecting it, and the fastest-growing revenue claims in physical AI belong to the cost of teaching.

## Sources

- [Mecka, Series B announcement](https://www.mecka.ai/news/series-b)
- [TechCrunch, Mecka raises $60M from Sequoia](https://techcrunch.com/2026/10/07/robot-data-startup-mecka-ai-nabs-60m-from-sequoia/)
- [TechCrunch, XDOF in talks at a $1.2B valuation](https://techcrunch.com/2026/09/04/xdof-just-three-months-out-of-stealth-is-in-talks-for-a-series-b-at-a-1-2b-valuation/)
- [TechCrunch, AI labs paying XDOF for robot data](https://techcrunch.com/2026/06/17/collecting-robot-training-data-is-dirty-unglamorous-work-some-ai-labs-are-already-paying-xdof-to-do-it/)
- [XDOF, ABC-130K dataset](https://www.xdof.ai/blog/abc-130k)
- [Scale AI, physical AI data](https://scale.com/blog/physical-ai)
- [CNN, the people training humanoid robots](https://www.cnn.com/2026/04/04/tech/humanoid-robot-training-jobs-intl-hnk-dst)
- [Rest of World, Unitree's Shanghai IPO](https://restofworld.org/2026/unitree-china-humanoid-robot-shanghai-ipo/)
- [Humanoid Hub, IDC first-half humanoid shipments](https://www.humanoidhub.ai/news/agibot-leads-global-humanoid-robot-shipments-in-h1-2026-idc-data-shows)
- [Euronews, Denmark tests humanoids in supermarkets](https://www.euronews.com/video/2026/10/11/denmark-tests-humanoid-robots-in-supermarkets-to-help-to-ease-staff-workload-at-night)
- [Roboterminal, Figure's Index training loop](https://roboterminal.app/posts/figure-index-nscale-training-loop)
- [Roboterminal, open robot models download test](https://roboterminal.app/posts/open-robot-models-download-test)

## Citation

```
"{quote}"
Source: Roboterminal · {url}
```

---

Source: Roboterminal · https://roboterminal.app/posts/robot-data-sellers-vs-unitree.md