Asimov Teardown — YC W26 Humanoid Training Data Marketplace

TL;DR

Asimov is a Mercor for humanoid robot data. The Mercor parallel is exact: AI lab pays, contractor performs, marketplace clips margin. Swap "labeled text from a Stanford grad" for "video of a line cook folding dough" and the unit economics rhyme.

The pitch is unglamorous in the best way. Frontier robotics labs — Figure AI, 1X, Tesla Optimus, Apptronik — need millions of hours of human-task footage to train their humanoid models. Synthetic data plateaus. Internal teleoperation studios cost $400/hour and don't scale. So Asimov runs the supply side: 5,000+ contributors with Asimov-provided hardware, embedded in households, restaurants, hotels, and factories, capturing the long tail of human movement that lab employees in San Francisco cannot.

Copyable score:

• Capital: 15/100 — hardware deployment to 5,000 nodes is not a weekend project. You need ~$2-5M to clone this with any credibility.
• Stack: 30/100 — cameras, IMUs, cloud ingest, labeling — all off-the-shelf. Integration is the moat.
• Channel: 25/100 — getting Figure AI to sign a data contract requires warm intros. YC W26 partners did this; you cannot.
• Network: 20/100 — both sides are oligopolies. 4 real buyers. Contributor side is recruitable but churns hard.
• Timing: 80/100 — humanoid robot capex is in the steepest part of the curve. This window is open for 18-24 months.

Verdict: cool. Not because the business is bad — it's probably excellent — but because the buyer set is closed, the capital wall is real, and the YC channel is non-replicable. If you are reading this and thinking "I'll build this too," you are competing with a company that has Garry Tan's phone number and warm intros to Brett Adcock. The transferable lesson lives in the adjacent labeled-data marketplaces, not in cloning humanoid data itself. We get to that in the playbook.

5-Minute Walkthrough

I signed up as a contributor. Here's what actually happened.

The landing page is sparse — three sentences, an email field, and a "Join the network" button. No pricing. No demo. No "Trusted by Figure AI" logo wall, which is actually a signal that they probably are working with Figure AI and the NDA precludes the brag. The signup flow asks for location, primary occupation, and a free-text field asking what tasks I perform on a typical day. The form is short enough that I finished it during one paragraph of a podcast.

Then nothing. No instant onboarding email. No Calendly. A 48-hour silence followed by a real human reply asking if I'd be open to a 15-minute video screening call.

This is the first interesting detail. Asimov is not a self-serve marketplace. It is a curated marketplace where the supply side is screened. The reason becomes obvious when you think about it: a lab paying $X per hour for "human folding laundry" footage cannot tolerate adversarial submissions, edge-case background mess, or contributors who clip and resell footage to a competitor. The screening call is a moat disguised as friction.

I declined the call — I am not actually going to embed an Asimov camera in my kitchen for a teardown — but I asked the operator a few questions over email. They were guarded but answered three:

1. Hardware is provided. You don't buy anything.
2. Payment is per accepted hour of footage, with a quality multiplier.
3. Contributors retain no rights to the footage once submitted.

That last point is the whole business. The contributor is selling not their time but the ownership of a recorded moment of their life. The marketplace's job is to make that trade feel normal. It does, mostly, because the contributor never sees what the lab does with t