Mecka AI is making $100M a year sending humans to collect robot training data
The bottleneck in physical AI isn't hardware anymore. It's data. To be fair, hardware is still hard, but data to drive better physical AI is seemingly harder, at least to scale.
Mecka AI has figured out a way to fill that gap, and it's generating $100 million a year doing it. Sequoia Capital just led a round that values the company near $500 million, with NVIDIA, Microsoft M12, and Qualcomm also participating. The Bloomberg report puts this as one of the most significant infrastructure bets in the robotics sector this year.
The model is straightforward: Mecka pays crowdsourced workers to perform tasks — making coffee, doing automotive repair, assembling parts — while wearing body sensors and carrying iPhones. The resulting motion capture data gets packaged and sold to robotics companies that need their models to understand how humans move through physical space.
Why this is harder than it sounds
Language models learned from the internet, but humanoid robots can't. There's no massive global archive of documented physical manipulation to train on. Basically every robot company building a general-purpose model needs original data: real humans performing real tasks in real environments, with the motion, force, and spatial relationships captured at sufficient resolution to be useful.
Getting that data at scale requires infrastructure. You need workers, sensors, protocols, quality control, and a distribution pipeline. Mecka has built all of it. Their clients include humanoid companies like 1X Technologies, which needs training data to improve its EVE and NEO robots' physical dexterity.
The $100M annual run rate is the number that should land hardest. That's current revenue. It means robot companies are already paying, at scale, for human movement data.
Crowdsourcing physical AI
Mecka's approach is structurally similar to how companies like Scale AI built the annotation infrastructure for computer vision and language models: make humans the raw data source, build quality control and tooling around them, and sell the output to AI developers.
The difference is that Scale workers label images and text. Mecka workers move their bodies.
The founders — CEO Josh Gao, CTO Jason Chong, COO Duy Nguyen, CPO Mogen Cheng — come from fintech and crypto, not robotics. That kinda makes sense ... building this kind of data marketplace is a logistics and operations problem as much as a technical one.
The company raised $8 million in August 2025, $25 million in a Series A in November 2025, then a further $35 million. The Sequoia-led Series B at near-$500M valuation represents a significant step change: Sequoia doesn't price rounds at half a billion for companies without strong conviction in their trajectory.
What this means for the humanoid industry
Every major humanoid company is, in theory, a competitor to Mecka: they could collect their own data rather than buying it. Figure, Agility, and Boston Dynamics all have internal data flywheel strategies. But internal data collection has limits: it's expensive, can be slow to scale, and it can be biased toward the tasks you already know you need.
(Figure AI, of course, is solving the last 2 of those with Index.)
A third-party data marketplace solves a different problem. It gives smaller humanoid companies access to diverse training data they couldn't generate themselves. It lets companies quickly add coverage for new task categories without deploying their own robot fleet to collect it. And at $100M ARR, Mecka has clearly found enough buyers to sustain that positioning.
The NVIDIA and Qualcomm participation is also interesting. Both are chip companies that benefit when humanoid robots consume more compute for both training and inference.
Backing the data infrastructure layer is a way to accelerate the market they're selling into.
The data layer becomes its own business
Training data as a standalone business was counterintuitive five years ago. Scale AI made it obvious for language and vision. Mecka is making the same case for physical AI.
At $300 million projected revenue for 2026, this is no longer a niche play. Mecka has identified a structural need — robot companies will always need more diverse, higher-quality human motion data than they can self-generate — and built infrastructure around it.
Sequoia's bet is that the need grows with the industry.
Given that the humanoid industry is projecting millions of deployed units by 2030, the amount of training data required to reach that scale is genuinely staggering. Someone has to collect it. Mecka is the clearest sign we have that "someone" might be a dedicated, independent business.
Sources
Written by
I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”