Data Marketplaces & Intermediaries

Realset AI raises $10M on a blunt premise: the physical world can't be scraped

Source: PR Newswire · Sep 29, 2026

Realset AI, a San Jose real-world data lab, has raised a $10 million Series A alongside Flatkey to build training data for frontier models and embodied agents. Its method is deliberately physical: egocentric video capture of skilled workers, reinforcement-learning environments reconstructed from real workflows, and expert evaluation of how deployed agents actually perform.

What makes the raise notable is the thesis behind it. Founder Hunter Guo's argument is blunt — the easy, scrapeable web is exhausted, and the next increment of model quality has to come from captured real-world human performance, because the physical world simply can't be scraped. That is close to verbatim the premise we've been building on: once the open web is mined out, the scarce input becomes authentic records of real humans doing real things, captured with consent and provenance intact.

For a structured interview corpus, Realset is both validation and a useful contrast. It validates the core claim that captured real-human performance is where value is migrating. The contrast is domain: Realset captures physical/skilled-work demonstrations, while an interview corpus captures the reasoning, probing, and evaluation of expert conversation — a different, equally un-scrapeable slice of real human behavior. Both point the same direction: the companies that own clean, consented real-world data will own the input the next generation of models can't synthesize.

Key Points

• San Jose real-world data lab raised a $10M Series A with Flatkey to build training data for frontier and embodied-agent models
• Captures egocentric video of skilled workers, builds RL environments from real workflows, and runs expert evaluation of deployed agents
• Founder Hunter Guo's framing: the scrapeable web is exhausted; value now lives in captured real-world human performance
• Almost verbatim the multimodal-interview-corpus thesis — real, captured human performance as the scarce input