Sequoia Capital is leading a funding round that could value Mecka AI at $500 million. The two-year-old startup pays people to record themselves doing ordinary things — making coffee, repairing a car — using body-worn sensors and smartphones, and sells the recordings as training data for robots. The new round extends financing Mecka AI announced three months ago, a $60 million deal that also included Menlo Ventures, SV Angel and Kindred Ventures. Mecka AI did not respond to a request for comment. Sequoia Capital declined to comment.
The interval is the number to look at. Three months separate a $60 million round from a round priced at half a billion dollars. Nothing in the company's public record explains that on operating grounds; what changed in between is the market's belief about robot training data, not anything Mecka AI has disclosed about its own business.
The one operating figure available is a projection. In early June, co-founder Josh Gao told Fortune the company expects to finish 2026 at a $100 million annual revenue run rate — meaning current monthly performance, annualised, would reach that figure. A $500 million valuation against that is five times a revenue number the founder expects to hit at the end of a year that still has three months to run. Investors are not pricing revenue. They are pricing a forecast of revenue, in a category where nobody has yet shown what the data is worth to the models trained on it.
The founding story is unusual enough to be worth stating plainly. Four entrepreneurs started the company in 2024. Josh Gao and Mogen Cheng, both Canadian, previously built a fintech startup for restaurants. Jason Chong joined Coinbase after it acquired his cryptocurrency exchange. Duy Nguyen, the only non-Canadian among them, runs operations. None of the four had worked in robotics. They noticed a shortage of physical-world data and concluded that capturing real-environment interactions had become the main obstacle to general-purpose robots, humanoids included.
The name comes from "mecha", the fictional giant robot piloted by a human — which is close to a description of the product, since a Mecka contractor with sensors strapped on is a human being instrumented so a machine can learn the movement. The commercial model is borrowed, openly, from Scale AI, Mercor, Surge and the other human-data suppliers that grew up around large language models.
That borrowed model is also the risk. The human-data business has already run its course once: margins compress as collection commoditises, the largest buyer eventually decides to collect in-house, and the biggest supplier gets absorbed by its biggest customer. Mecka AI is running that playbook again, one layer down the stack, with a data type that is harder to gather and — this is the part the round does not settle — not obviously harder to gather yourself. Egocentric recordings sit alongside teleoperation in most labs' pipelines, not above them.
Notably absent from anything the company has said: a single customer name. For a business projecting a $100 million run rate, that omission is louder than the valuation. A named lab under contract would tell the market that frontier robotics teams have decided to buy this data rather than capture it. Without one, the $500 million rests on the premise that they will.
The comparison set says the premise is widely held. XDOF, another real-world robot data startup, is approaching a round at a $1.2 billion valuation, TechCrunch reported last week. Scale AI and Micro1 are both pushing their human-data platforms past language models. Three rounds and two incumbents are moving on the same thesis at once, which is what a consensus looks like before it has been tested by a customer.
Robot data is being priced the way language data was in 2021, by firms that believe the bottleneck will hold. The bottleneck holds only as long as robots stay out of the world. Once machines are deployed and running, they record everything they do, for free, forever — and the companies buying Mecka AI's footage today are the ones building them.